Abuse-Resistant OpenPGP Keystores

The key words “MUST”, “MUST NOT”, “REQUIRED”, “SHALL”, “SHALL NOT”, “SHOULD”, “SHOULD NOT”, “RECOMMENDED”, “NOT RECOMMENDED”, “MAY”, and “OPTIONAL” in this document are to be interpreted as described in BCP 14 when, and only when, they appear in all capitals, as shown here.

“OpenPGP certificate” (or just “certificate”) is used interchangeably with ’s “Transferable Public Key”. The term “certificate” refers unambiguously to the entire composite object, unlike “key”, which might also be used to refer to a primary key or subkey. An “identity certification” (or just “certification”) is an signature packet that covers OpenPGP identity information – that is, any signature packet of type 0x10, 0x11, 0x12, or 0x13. Certifications are said to (try to) “bind” a primary key to a User ID. The primary key that makes the certification is known as the “issuer”. The primary key over which the certification is made is known as the “subject”. A “first-party certification” is issued by the primary key of a certificate, and binds itself to a user ID in the certificate. That is, the issuer is the same as the subject. This is sometimes referred to as a “self-sig”. A “third-party certification” is a made over a primary key and user ID by some other certification-capable primary key. That is, the issuer is different than the subject. (The elusive “second-party” is presumed to be the verifier who is trying to interpret the certificate) All subkeys are bound to the primary key with an Subkey Binding Signature. Some subkeys also reciprocate by binding themselves back to the primary key with an Primary Key Binding Signature. The Primary Key Binding Signature is also known as a “cross-signature” or “cross-sig”. A “keystore” is any collection of OpenPGP certificates. Keystores typically receive mergeable updates over the course of their lifetime which might add to the set of OpenPGP certificates they hold, or update the certificates. “Certificate validation” is the process whereby a user decides whether a given user ID in an OpenPGP certificate is acceptable for use. For example, if the certificate has a user ID of Alice <alice@example.org> and the user wants to send an e-mail to alice@example.org, the mail user agent might want to ensure that the certificate is valid for this e-mail address before encrypting to it. Some clients may rely on specific keystores for certificate validation, but some keystores (e.g., ) make no assertions whatsoever about certificate validity, and others offer only very subtle guarantees. See for more details. “Certificate lookup” refers to the retrieval of a set of certificates from a keystore based on the user ID or some substring match of the user ID. See for more details. “Certificate update” refers to retrieval of a certificate from a keystore based on the fingerprint of the primary key. See for more details. “Certificate discovery” refers to the retrieval of a set of certificates from a keystore based on the fingerprint or key ID of any key in the certificate. See for more details. A “keyserver” is a particular kind of keystore, typically a means of publicly distributing OpenPGP certificates or updates to them. Examples of keyserver software include and . One common HTTP interface for keyservers is . A “synchronizing keyserver” is a keyserver which gossips with other peers, and typically acts as an append-only log. Such a keyserver is typically useful for certificate lookup, certificate discovery, and certificate update (including revocation information). They are typically not useful for certificate validation, since they make no assertions about whether the identities in the certificates they server are accurate. As of the writing of this document, is the canonical synchronizing keyserver implementation, though other implementations exist. An “e-mail validating keyserver” is a keyserver which attempts to verify the identity in an OpenPGP certificate’s user ID by confirming access to the e-mail account, and possibly by confirming access to the secret key. Some implementations permit removal of a certificate by anyone who can prove access to the e-mail address in question. They are useful for certificate lookup based on e-mail address and certificate validation (by users who trust the operator), but some may not be useful for certificate update or certificate discovery, since a certificate could be simply replaced by an adversary who also has access to the e-mail address in question. is an example of such a keyserver. “Cryptographic validity” refers to mathematical evidence that a signature came from the secret key associated with the public key it claims to come from. Note that a certification may be cryptographically valid without the signed data being true (for example, a given certificate with the user ID Alice <alice@example.org> might not belong to the person who controls the e-mail address alice@example.org even though the self-sig is cryptographically valid). In particular, cryptographic validity for user ID in a certificate is typically insufficient evidence for certificate validation. Also note that knowledge of the public key of the issuer is necessary to determine whether any given signature is cryptographically valid. Some keyservers perform cryptographic validation in some contexts. Other keyservers (like ) perform no cryptographic validation whatsoever. OpenPGP revocations can have “Reason for Revocation” (see ), which can be either “soft” or “hard”. The set of “soft” reasons is: “Key is superseded” and “Key is retired and no longer used”. All other reasons (and revocations that do not state a reason) are “hard” revocations. See for more detail.

OpenPGP keystores that handle submissions from the public are subject to a range of attacks by malicious submitters. This section describes five distinct attacks that public keystores should consider. The rest of the document describes some mitigations that can be used by keystores that are concerned about these problems but want to continue to offer some level of service for certificate lookup, certificate update, certificate discovery, or certificate validation.

Many public keystores (including both the keyserver network and ) allow anyone to attach arbitrary data (in the form of third-party certifications) to any certificate, bloating that certificate to the point of being impossible to effectively retrieve. For example, some OpenPGP implementations simply refuse to process certificates larger than a certain size. This kind of Denial-of-Service attack makes it possible to make someone else’s certificate unretrievable from the keystore, preventing certificate lookup, discovery, or update. In the case of a revoked certificate that has been flooded, this potentially leaves the client of the keystore with the compromised certificate in an unrevoked state locally because it was unable to fetch the revocation information. Additionally, even without malice, OpenPGP certificates can potentially grow without bound.

Public keystores that are used for certificate lookup may also be vulnerable to attacks that flood the space of known user IDs. In particular, if the keystore accepts arbitrary certificates from the public and does no verification of the user IDs, then any client searching for a given user ID may need to review and process an effectively unbounded set of maliciously-submitted certificates to find the non-malicious certificates they are looking for. For example, if an attacker knows that a given system consults a keystore looking for certificates which match the e-mail address alice@example.org, the attacker may upload hundreds or thousands of certificates containing user IDs that match that address. Even if those certificates would not be accepted by a client (e.g., because they were not certified by a known-good authority), the client typically still has to wade through all of them in order to find the non-malicious certificates. If the keystore does not offer a lookup interface at all (that is, if clients cannot search it by user ID), then user ID flooding is of less consequence.

A malicious actor who wants to render a certificate unavailable for update may generate an arbitrary number of OpenPGP certificates with the targeted primary key attached as a subkey. If they can convince a keystore to accept all of those certificates, and the keystore returns them by subkey match during certificate update, then the certificate update client will need to spend an arbitrary amount of bandwidth and processing power filtering out the irrelevant data, and may potentially give up before discovering the certificate of interest. A malicious actor may also want to confuse a certificate discovery request that was targeted at a particular subkey, by binding that subkey to multiple bogus certificates. If these bogus certificates are ingested and redistributed by the keystore, then a certificate discovery client may receive a set of certificates that cannot be adequately distinguished.

A public keystore that accepts arbitrary OpenPGP material and is append-only is at risk of being overwhelmed by sheer quantity of malicious uploaded packets. This is a risk even if the user ID space is not being deliberately flooded, and if individual certificates are protected from flooding by any of the mechanisms described later in this document. The keystore itself can become difficult to operate if the total quantity of data is too large, and if it is a synchronizing keyserver, then the quantities of data may impose unsustainable bandwidth costs on the operator as well. Effectively mitigating against keystore flooding requires either abandoning the append-only property that some keystores prefer, or imposing very strict controls on initial ingestion.

Like any large public dataset, it’s possible that a keystore ends up hosting some content that is legally actionable in some jurisdictions, including libel, child pornography, material under copyright or other “intellectual property” controls, blasphemy, hate speech, etc. A public keystore that accepts and redistributes arbitrary content may face risk due to uploads of toxic data.

Some keystores have simple interfaces, like files present in a local filesystem. But many keystores offer an API for certificate retrieval of different types. This section documents a set of useful interactions that a client may have with such a keystore. They are represented in abstract form, and are not intended to be the full set of interfaces offered by any keystore, but rather a convenient way to think about the operations that make the keystore useful for its clients. Not all keystores may offer all of these interfaces, or they may offer them in subtly different forms, but clients will nevertheless try to perform something like these operations with keystores that they interact with.

This is the simplest keystore operation. The client sends the keystore the full fingerprint of the certificate’s primary key, and the keystore sends the client the corresponding certificate (or nothing, if the keystore does not contain a certificate with a matching primary key).

certificate? ]]> A client uses certificate update to retrieve the full details of a certificate that it already knows about. For example, it might be interested in updates to the certificate known to the keystore, including revocations, expiration updates, new third-party certifications, etc. Upon successful update, the client SHOULD merge the retrieved certificate with its local copy. Not all keystores offer this operation. For example, clients cannot use WKD () or OPENPGPKEY ( for certificate update.

If a client is aware of an OpenPGP signature or certification that it cannot verify because it does not know the issuing certificate, it may consult a keystore to try to discover the certificate based on the Issuer or Issuer Fingerprint subpacket in the signature or certification it is trying to validate.

certificate_list ]]> This is subtly different from certificate update () in three ways: it may return more than one certificate (e.g., when multiple certificates share a subkey, or when a primary key on one certificate is a subkey on another) it is willing to accept searches by short key ID, not just fingerprint it is willing to match against a subkey, not just a primary key While a certificate discovery client does not initially know the certificate it is looking for, it’s possible that the returned certificate is one that the client already knows about. For example, a new subkey may have been added to a certificate. Upon successful discovery, the client SHOULD merge any retrieved certificates with discovered local copies (as determined by primary key), and then evaluate the original signature against any retrieved certificate that appears to be valid and reasonable for use in the signing context. It is unclear what a client should do if multiple certificates do appear to be valid for a given signature, because of ambiguity this represents about the identity of the signer. However, this ambiguity is similar to the ambiguity of a certificate with multiple valid user IDs, which the client already needs to deal with. Not all keystores offer this operation. For example, clients cannot use WKD () or OPENPGPKEY ( for certificate discovery.

If a client wants to encrypt a message to a particular e-mail address, or wants to encrypt a backup to some identity that it knows of but does not have a certificate for, it may consult a keystore to discover certificates that claim that identity in their user ID packets. Both and offer certificate lookup mechanisms. User IDs are constrained only in that they are a UTF-8 string, but some conventions govern their practical use. See for more discussion of some common conventions around user ID structure. Note that lookup does not necessarily imply user ID or certificate validation. It is entirely possible for a keystore to return a certificate during lookup that the client cannot validate. Abuse-resistant keystores that offer a lookup interface SHOULD distinguish interfaces that perform full-string-match lookup from interfaces that perform e-mail address based lookup.

The most straightforward form of certificate lookup asks for the set of all certificates that contain a user ID that exactly and completely matches the query parameter supplied by the client.

certificate_list ]]> In its simplest form, this match is done by a simple bytestring comparison. More sophisticated keystores MAY perform the comparison after applying form NFC to both the uid query and the user IDs from the stored certificates.

However, some common use cases look for specific patterns in the user ID rather than the entire user ID. Most useful to many existing OpenPGP clients is a lookup by e-mail address.

certificate_list ]]> For certificates with a user ID that matches the structure of an name-addr or addr-spec, a keystore SHOULD extract the addr-spec from the user ID, canonicalize it (see ), and compare it to the canonicalized form of of the addr query parameter.

Some keystores offer other forms of substring or regular expression matching against the stored user IDs. These other forms of lookup may be useful in some contexts (e.g., ), but they may also represent privacy concerns (e.g., ), and they may impose additional computational or indexing burdens on the keystore.

An OpenPGP client may assess certificate and user ID validity based on many factors, some of which are directly contained in the certificate itself (e.g., third-party certifications), and some of which are based on the context known to the client, including: Whether it has seen e-mails from that address signed by that certificate in the past, How long it has known about the certificate, Whether the certificate was fetched from a keystore that asserts validity of the user ID or some part of it (such as the e-mail address). A keystore MAY facilitate clients pursuing this last point of contextual corroboration via a direct interface:

boolean ]]> In an e-mail-specific context, the client might only care about the keystore’s opinion about the validity of the certificate for the e-mail address portion of the user ID only:

boolean ]]> For some keystores, the presence of a certificate in the keystore alone implies that the keystore asserts the validity of all user IDs in the certificate retrieved. For others, the presence in the keystore applies only to some part of the user ID. For example, will only return user IDs that have completed an e-mail validation step, so presence in that keystore implies an assertion of validity of the e-mail address part of the user IDs returned, but makes no claim about the display-name portion of any returned user IDs. Note that a client retrieving a certificate from such a keystore may merge the certificate with a local copy – but the validity asserted by the keystore of course has no bearing on the packets that the keystore did not return. In a more subtle example, the retrieval of a certificate looked up via WKD () or DANE () should only be interpreted as a claim of validity about any user ID which matches the e-mail address by which the certificate was looked up, with no claims made about any display-name portions, or about any user ID that doesn’t match the queried e-mail address at all. A keystore that offers some sort of validation interface may also change its opinion about the validity of a given certificate or user ID over time; the interface described above only allows the client to ask about the keystore’s current opinion, but a more complex interface might be capable of describing the keystore’s assertion over time. See also . An abuse-resistant keystore that clients rely on for any part of their certificate validation process SHOULD offer a distinct interface for making assertions about certificate and user ID validity to help clients avoid some of the subtleties involved with inference based on presence described above. Note that the certificate validation operation as described above has a boolean response. While a “true” response indicates that keystore believes the user ID or e-mail address is acceptable for use with the certificate referred to by the public key fingerprint, a “false” response doesn’t necessarily mean that the keystore actively thinks that the certificate is actively bad, or must not be used for the referenced identity. Rather, “false” is the default state: no opinion is expressed by the keystore, and the client is left to make their own inference about validity based on other factors. A keystore MAY offer a more nuanced validity interface; if it does, it SHOULD explicitly document the semantics of the different response types so that clients can make appropriate judgement.

Different keystores have different ways to submit a certificate for consideration for ingestion, including: a simple upload of a certificate via http round-trip e-mail verification proof of presence in some other service vouching, or other forms of multi-party attestation Because these schemes vary so widely, this document does not attempt to describe the keystore certificate submission process in detail. However, guidance can be found for implementations that generate, manage, and submit certificates in .

These steps can be taken by any keystore that wants to avoid obviously malicious abuse. They can be implemented on receipt of any new packet, and are based strictly on the structure of the packet itself.

While permits OpenPGP packet sizes of arbitrary length, OpenPGP certificates rarely need to be so large. An abuse-resistant keystore SHOULD reject any OpenPGP packet larger than 8383 octets. (This cutoff is chosen because it guarantees that the packet size can be represented as a one- or two-octet “New Format Packet Length”, but it could be reduced further) This may cause problems for user attribute packets that contain large images, but it’s not clear that these images are concretely useful in any context. Some keystores MAY extend this limit for user attribute packets specifically, but SHOULD NOT allow even user attributes packets larger than 65536 octets.

indicates that User IDs are expected to be UTF-8 strings. An abuse-resistant keystore MUST reject any user ID that is not valid UTF-8. Some abuse-resistant keystores MAY only accept User IDs that meet even stricter conventions, such as an name-addr or addr-spec, or a URL like ssh://host.example.org (see ). As simple text strings, User IDs don’t need to be nearly as long as any other packets. An abuse-resistant keystore SHOULD reject any user ID packet larger than 1024 octets.

Some abuse-resistant keystores may restrict themselves to publishing only certificates with User IDs that match a specific pattern. For example, encourages publication in the DNS of only certificates whose user IDs refer to e-mail addresses within the DNS zone. similarly aims to restrict publication to certificates relevant to the specific e-mail domain.

signature packets contain an “unhashed” block of subpackets. These subpackets are not covered by any cryptographic signature, so they are ripe for abuse. An abuse-resistant keystore SHOULD strip out all unhashed subpackets but the following exceptions:

Some certifications only identify the issuer of the certification by an unhashed Issuer or Issuer Fingerprint subpacket. If a certification’s hashed subpacket section has no Issuer Fingerprint (see ) subpacket, then an abuse-resistant keystore that has cryptographically validated the certification SHOULD synthesize an appropriate Issuer Fingerprint subpacket and include it in the certification’s unhashed subpackets.

Some Primary Key Binding Signatures (“cross-sigs”) are distributed as unhashed subpackets in a Subkey Binding Signature. A cryptographically-validating abuse-resistant keystore SHOULD be willing to redistribute a valid cross-sig as an unhashed subpacket. The redistributed unhashed cross-sig itself should be stripped of all unhashed subpackets.

Some third-party certifications are attested to by the certificate primary key itself in an unhashed subpacket, as described in . A cryptographically-validating abuse-resistant keystore SHOULD be willing to redistribute a valid first-party attestation as an unhashed subpacket. The redistributed first-party attestation itself should be stripped of all unhashed subpackets.

Due to size concerns, some abuse-resistant keystores MAY choose to ignore user attribute packets entirely, as well as any certifications that cover them.

An abuse-resistant keystore MUST NOT accept any certification that has the “Exportable Certification” subpacket present and set to 0. While most keystore clients will not upload these “local” certifications anyway, a reasonable public keystore that wants to minimize data has no business storing or distributing these certifications.

Many OpenPGP packets have time-of-creation timestamps in them. An abuse-resistant keystore with a functional real-time clock MAY decide to only accept packets whose time-of-creation is in the past. Note that some OpenPGP implementations may pre-generate OpenPGP material intended for use only in some future window (e.g. “Here is the certificate we plan to use to sign our software next year; do not accept signatures from it until then.”), and may use modified time-of-creation timestamps to try to achieve that purpose. This material would not be distributable ahead of time by an abuse-resistant keystore that adopts this mitigation.

An aggressively abuse-resistant keystore MAY decide to only accept certifications that meet a specific profile. For example, it MAY reject certifications with unknown subpacket types, unknown notations, or certain combinations of subpackets. This can help to minimize the amount of room for garbage data uploads. Any abuse-resistant keystore that adopts such a strict posture should clearly document what its expected certificate profile is, and should have a plan for how to extend the profile if new types of certification appear that it wants to be able to distribute. Note that if the profile is ever restricted (rather than extended), and the restriction is applied to the material already present, such a keystore is no longer append-only (please see ).

An abuse-resistant keystore capable of cryptographic validation MAY retain a list of designated authorities, typically in the form of a set of known public keys. Upon receipt of a new OpenPGP certificate, the keystore can decide whether to accept or decline each user ID of the certificate based whether that user ID has a certification that was issued by one or more of the designated authorities. If no user IDs are certified by designated authority, such a keystore SHOULD decline the certificate and its primary key entirely. Such a keystore SHOULD decline to retain or propagate all certifications associated with each accepted user ID except for first-party certifications and certifications by the designated authorities. The operator of such a keystore SHOULD have a clear policy about its set of designated authorities. Given the ambiguities about expiration and revocation, such a keyserver SHOULD ignore expiration and revocation of authority certifications, and simply accept and retain as long as the cryptographic signature is valid. Note that if any key is removed from the set of designated authorities, and that change is applied to the existing keystore, such a keystore may no longer be append-only (please see ).

The maintainer of the keystore may keep a specific list of “known-bad” material, and decline to accept or redistribute items matching that blocklist. The material so identified could be anything, but most usefully, specific public keys or User IDs could be blocked. Note that if a blocklist grows to include an element already present in the keystore, it will no longer be append-only (please see ). Some keystores may choose to apply a blocklist only at retrieval time and not apply it at ingestion time. This allows the keystore to be append-only, and permits synchronization between keystores that don’t share a blocklist, and somewhat reduces the attacker’s incentive for flooding the keystore (see for more discussion). Note that development and maintenance of a blocklist is not without its own potentials for abuse. For one thing, the blocklist may itself grow without bound. Additionally, a blocklist may be socially or politically contentious as it may describe data that is toxic () in one community or jurisdiction but not another. There needs to be a clear policy about how it is managed, whether by delegation to specific decision-makers, or explicit tests. Furthermore, the existence of even a well-intentioned blocklist may be an “attractive nuisance,” drawing the interest of would-be censors or other attacker interested in controlling the ecosystem reliant on the keystore in question.

Most of the abuse mitigations described in this document are described as being applied at certificate ingestion time. It’s also possible to apply the same mitigations when a certificate is retrieved from the keystore (that is, during certificate lookup, update, or discovery). Applying an abuse mitigation at retrieval time may help a client defend against a user ID flooding (), certificate flooding (), or fingerprint flooding () attack. It may also help a keystore limit its liability for redistributing toxic data (). However, only mitigations applied at ingestion time are able to mitigate keystore flooding attacks (). Some mitigations (like the non-append-only mitigations described in ) may be applied as filters at retrieval time, while still allowing access to the (potentially much larger) unfiltered dataset associated given certificate or user ID via a distinct interface. The rest of this section documents specific mitigations that are only relevant at retrieval time (certificate discovery, lookup, or update).

Some abuse-resistant keystores may accept and store user IDs but decline to redistribute some or all of them, while still distributing the certifications that cover those redacted user IDs. This draft refers to such a keystore as a “user ID redacting” keystore. The certificates distributed by such a keystore are technically invalid “transferable public keys”, because they lack a user ID packet, and the distributed certifications cannot be cryptographically validated independently. However, an OpenPGP implementation that already knows the user IDs associated with a given primary key will be capable of associating each certification with the correct user ID by trial signature verification.

A user ID redacting keystore is useful for certificate update by a client that already knows the user ID it expects to see associated with the certificate. For example, a client that knows a given certificate currently has two specific user IDs could access the keystore to learn that one of the user IDs has been revoked, without any other client learning the user IDs directly from the keystore.

A user ID redacting keystore is somewhat less useful for clients doing certificate discovery. Consider the circumstance of receiving a signed e-mail without access to the signing certificate. If the verifier retrieves the certificate from a user ID redacting keystore by via the Issuer Fingerprint from the signature, and the signature validates, the received certificate might not be a valid “transferable public key” unless the client can synthesize the proper user ID. A reasonable client that wants to validate a certification in the user ID redacted certificate SHOULD try to synthesize possible user IDs based on the value of the From: header in the message: Decode any encodings present in the raw header value, converting into UTF-8 form C (NFC), trimming all whitespace from the beginning and the end of the string. The resulting string should be an name-addr or addr-spec. If it is a name-addr, convert the UTF-8 string into an OpenPGP user ID and check whether the certification validates, terminating on success. If the test fails, extract the addr-spec from the name-addr and continue. Canonicalize the addr-spec according to , and check whether the certification validates, terminating on success. If it doesn’t validate wrap the canonicalized addr-spec in angle-brackets (“<” and “>”) and test the resulting minimalist name-addr against the certification, terminating on success. If all of the above fails, synthesis has failed.

It’s possible (though non-intuitive) to use a user ID redacting keystore for certificate lookup. Since the keystore retains (but does not distribute) the user IDs, they can be used to select certificates in response to a search. The OpenPGP certificates sent back in response to the search will not contain the user IDs, but a client that knows the full user ID they are searching for will be able to verify the returned certifications. Certificate lookup from a user ID redacting keystore works better for certificate lookup by exact user ID match than it does for substring match, because a client that retrieves a certificate via a substring match may not be able to reconstruct the redacted user ID. However, without some additional restrictions on which certifications are redistributed (whether the user ID is redacted or not), certificate lookup can be flooded (see ).

To ensure that the distributed certificate is at least structurally a valid transferable public key, a user ID redacting keystore MAY distribute an empty user ID (an OpenPGP packet of tag 13 whose contents are a zero-octet string) in place of the omitted user ID. This two-octet replacement user ID packet (“\xb4\x00”) is called the “unstated user ID”. To facilitate clients that match certifications with specific user IDs, a user ID redacting keystore MAY insert a non-hashed notation subpacket into the certification. The notation will have a name of “uidhash”, with 0x80 (“human-readable”) flag unset. The value of such a notation MUST be 32 octets long, and contains the SHA-256 cryptographic digest of the UTF-8 string of the redacted user ID. A certificate update client which receives such a certification after the “unstated user ID” SHOULD compute the SHA-256 digest of all user IDs it knows about on the certificate, and compare the result with the contents of the “uidhash” notation to decide which user ID to try to validate the certification against.

User ID redaction is at best an imperfect process. Even if a keystore redacts a User ID, if it ships a certification over that user ID, an interested client can guess user IDs until it finds one that causes the signature to verify. This is even easier when the space of legitimate user IDs is relatively small, such as the set of commonly-used hostnames

Abuse-resistant keystores can defend against a fingerprint flooding attack during certificate update by implementing a narrowly-constrained certificate update interface. Such a keystore MUST accept only a full fingerprint as the search parameter from the certificate update client, and it MUST return at most a single certificate whose primary key matches the requested fingerprint. It MUST NOT return more than one certificate, and it MUST NOT return any certificate whose primary key does not match the fingerprint. In particular, it MUST NOT return certificates where only the subkey fingerprint matches. Note that does not offer the primitive described in exactly. In that specification, the set of keys returned by a “get” operation with a “search” parameter that appears to be a full fingerprint is ambiguous. Some popular implementations (e.g., ) do not currently implement this mitigation, because they return certificates with subkeys that match the fingerprint.

By definition, certificate discovery needs to be able to match subkeys, not just primary keys. This means that the mitigation in is ineffective for a keystore that offers a certificate discovery interface. An abuse-resistant keystore that aims to defend its certificate discovery interface from a fingerprint flooding () attack can follow the following procedure. Such a keystore MUST accept only a full fingerprint or a 64-bit key ID as the search parameter from the certificate discovery client. It MUST only match that fingerprint against the following: the fingerprint or key ID of a primary key associated with a valid certificate the fingerprint or key ID of a cryptographically-valid subkey that also has a cross-sig. This defends against the fingerprint flooding attack because a certificate will only be returned by subkey if the subkey has agreed to be associated with the primary key (and vice versa). Note that this mitigation means that certificate discovery will fail if used for subkeys that lack cross-sigs. In particular, this means that a client that tries to use the certificate discovery interface to retrieve a certificate based on its encryption-capable subkey (e.g., taking the key ID from a Public Key Encrypted Session Key (PKESK) packet) will have no success. This is an acceptable loss, since the key ID in a PKESK is typically unverifiable anyway.

Some mitigations make the acceptance or rejection of packets contingent on data that is already in the keystore or the keystore’s developing knowledge about the world. This means that, depending on the order that the keystore encounters the various material, or how it accesses or finds the material, the final set of material retained and distributed by the keystore might be different. While this isn’t necessarily bad, it may be a surprising property for some users of keystores.

An abuse-resistant keystore that is capable of doing cryptographic validation MAY decide to reject certifications that it cannot cryptographically validate. This may mean that the keystore rejects some packets while it is unaware of the public key of the issuer of the packet.

This is an extension of , but where the set of authorities is just the set of certificates already known to the keystore. An abuse-resistant keystore that adopts this strategy is effectively only crawling the reachable graph of OpenPGP certificates from some starting core. A keystore adopting the mitigation SHOULD have a clear documentation of the core of initial certificates it starts with, as this is effectively a policy decision. This mitigation measure may fail due to a compromise of any secret key that is associated with a primary key of a certificate already present in the keystore. Such a compromise permits an attacker to flood the rest of the network. In the event that such a compromised key is identified, it might be placed on a blocklist (see ). In particular, if a public key is added to a blocklist for a keystore implementing this mitigation, and it is removed from the keystore, then all certificates that were only “reachable” from the blocklisted certificate should also be simultaneously removed.

Some OpenPGP keystores accept material from the general public over the Internet. If an abuse-resistant keystore observes a flood of material submitted to the keystore from a given Internet address, it MAY choose to throttle submissions from that address. When receiving submissions over IPv6, such a keystore MAY choose to throttle entire nearby subnets, as a malicious IPv6 host is more likely to have multiple addresses. This requires that the keystore maintain state about recent submissions over time and address. It may also be problematic for users who appear to share an IP address from the vantage of the keystore, including those behind a NAT, using a VPN, or accessing the keystore via Tor.

Some public keystores resist abuse by explicitly filtering OpenPGP material based on a set of external processes. For example, adjudicates the contents of the “Debian keyring” keystore based on organizational procedure and manual inspection.

Some keystores resist abuse by declining any certificate until the user IDs have been verified by e-mail. When these “e-mail validating” keystores review a new certificate that has a user ID with an e-mail address in it, they send an e-mail to the associated address with a confirmation mechanism (e.g., a high-entropy HTTPS URL link) in it. In some cases, the e-mail itself is encrypted to an encryption-capable key found in the proposed certificate. If the keyholder triggers the confirmation mechanism, then the keystore accepts the certificate. Some e-mail validating keystores MAY choose to distribute certifications over all user IDs for any given certificate, but will redact (see ) those user IDs that have not been e-mail validated. describes some concerns held by a keystore operator using this approach. is another example.

The following mitigations may cause some previously-retained packets to be dropped after the keystore receives new information, or as time passes. This is entirely reasonable for some keystores, but it may be surprising for any keystore that expects to be append-only (for example, some keyserver synchronization techniques may expect this property to hold). Furthermore, keystores that drop old data (e.g., superseded certifications) may make it difficult or impossible for their users to reason about the validity of signatures that were made in the past. See for more considerations. Note also that many of these mitigations depend on cryptographic validation, so they’re typically contextual as well. A keystore that needs to be append-only, or which cannot perform cryptographic validation MAY omit these mitigations. Alternately, a keystore may omit these mitigations at certificate ingestion time, but apply these mitigations at retrieval time (during certificate update, discovery, or lookup), and offer a more verbose (non-mitigated) interface for auditors, as described in . Note that anticipates some of these suggestions with its “clean” subcommand, which is documented as:

An abuse-resistant keystore SHOULD drop all signature packets that are explicitly superseded. For example, there’s no reason to retain or distribute a self-sig by key K over User ID U from 2017 if the keystore have a cryptographically-valid self-sig over <K,U> from 2019. Note that this covers both certifications and signatures over subkeys, as both of these kinds of signature packets may be superseded. Getting this right requires a nuanced understanding of subtleties in related to timing and revocation.

If a signature packet is known to only be valid in the past, there is no reason to distribute it further. An abuse-resistant keystore with access to a functional real-time clock SHOULD drop all certifications and subkey signature packets with an expiration date in the past. Note that this assumes that the keystore and its clients all have roughly-synchronized clocks. If that is not the case, then there will be many other problems!

If enough signature packets are dropped, it’s possible that some of the things that those signature packets cover are no longer valid. An abuse-resistant keystore which has dropped all certifications that cover a User ID SHOULD also drop the User ID packet. Note that a User ID that becomes invalid due to revocation MUST NOT be dropped, because the User ID’s revocation signature itself remains valid, and needs to be distributed. A primary key with no User IDs and no subkeys and no revocations MAY itself also be removed from distribution, though note that the removal of a primary key may make it impossible to cryptographically validate other certifications held by the keystore.

If the primary key of a certificate is revoked via a direct key signature, an abuse-resistant keystore SHOULD drop all the rest of the associated data (user IDs, user attributes, and subkeys, and all attendant certifications and subkey signatures). This defends against an adversary who compromises a primary key and tries to flood the certificate to hide the revocation. Note that the direct key revocation signature MUST NOT be dropped. In the event that an abuse-resistant keystore is flooded with direct key revocation signatures, it should retain the hardest, earliest revocation (see also ). In particular, if any of the direct key revocation signatures is a “hard” revocation, the abuse-resistant keystore SHOULD retain the earliest such revocation signature (by signature creation date). Otherwise, the abuse-resistant keystore SHOULD retain the earliest “soft” direct key revocation signature it has seen. If either of the above date comparisons results in a tie between two revocation signatures of the same “hardness”, an abuse-resistant keystore SHOULD retain the signature that sorts earliest based on a binary string comparison of the direct key revocation signature packet itself.

In combination with some of the dropping mitigations above, a particularly aggressive abuse-resistant keystore MAY choose an implicit expiration date for all signature packets. For example, a signature packet that claims no expiration could be treated by the keystore as expiring 3 years after issuance. This would permit the keystore to eject old packets on a rolling basis. An abuse-resistant keystore that adopts this mitigation needs a policy for handling signature packets marked with an explicit expiration that is longer than implicit maximum. The two obvious strategies are: cap the packet’s expiration to the system’s implicit expiration date, or accept the explicit expiration date. Warning: Any implementation of this idea is pretty radical, and it’s not clear what it would do to an ecosystem that depends on such a keystore. It probably needs more thinking.

In addition to the mitigations above, some keystores may resist abuse by declining to accept any user IDs or certifications whatsoever. Such a keystore MUST be capable of cryptographic validation. It accepts primary key packets, cryptographically-valid direct-key signatures from a primary key over itself, subkeys and their cryptographically-validated binding signatures (and cross-sigs, where necessary). A client of an updates-only keystore cannot possibly use the keystore for certificate lookup, because there are no user IDs to match. And it is not particularly useful for certificate discovery, because the returned certificate would have no identity information. However, such a keystore can be used for certificate update, as it’s possible to ship revocations (which are direct key signatures), new subkeys, updates to subkey expiration, subkey revocation, and direct key signature-based certificate expiration updates. Note that many popular OpenPGP implementations do not implement direct primary key expiration mechanisms, relying instead on user ID expirations. These user ID expiration dates or other metadata associated with a self-certification will not be distributed by an updates-only keystore. Certificates shipped by an updates-only keystore are technically invalid “transferable public keys,” because they lack a user ID packet. However many OpenPGP implementations will accept such a certificate if they already know of a user ID for the certificate, because the composite certificate resulting from a merge will be a standards-compliant transferable public key.

Slightly more permissive than the updates-only keystore described in is a keystore that also permits user IDs and their self-sigs. A first-party-only keystore only accepts and distributes cryptographically-valid first-party certifications. Given a primary key that the keystore understands, it will only attach user IDs that have a valid self-sig, and will only accept and re-distribute subkeys that are also cryptographically valid (including requiring cross-sigs for signing-capable subkeys as recommended in ). This effectively avoids certificate flooding attacks, because the only party who can make a certificate overly large is the holder of the secret corresponding to the primary key itself. Note that a first-party-only keystore is still problematic for those people who rely on the keystore for retrieval of third-party certifications. attempts to address this lack.

It is possible to operate an first-party-only keystore that redistributes certifications while declining to redistribute user IDs (see ). This defends against concerns about publishing identifiable information, while enabling full certificate update for those keystore clients that already know the associated user IDs for a given certificate.

We can augment a first-party-only keystore to allow it to distribute third-party certifications as long as the first-party has signed off on the specific third-party certification. An abuse-resistant keystore SHOULD only accept a third-party certification if it meets the following criteria: The third-party certification MUST be cryptographically valid. Note that this means that the keystore needs to know the primary key for the issuer of the third-party certification. The third-party certification MUST have an unhashed subpacket of type Embedded Signature, the contents of which we’ll call the “attestation”. This attestation is from the certificate’s primary key over the third-party certification itself, as detailed in the steps below: The attestation MUST be an OpenPGP signature packet of type 0x50 (Third-Party Confirmation signature) The attestation MUST contain a hashed “Issuer Fingerprint” subpacket with the fingerprint of the primary key of the certificate in question. The attestation MUST NOT be marked as non-exportable. The attestation MUST contain a hashed Notation subpacket with the name “ksok”, and an empty (0-octet) value. The attestation MUST contain a hashed “Signature Target” subpacket with “public-key algorithm” that matches the public-key algorithm of the third-party certification. The attestation’s hashed “Signature Target” subpacket MUST use a reasonably strong hash algorithm (as of this writing, any hash algorithm except MD5, SHA1, or RIPEMD160), and MUST have a hash value equal to the hash over the third-party certification with all unhashed subpackets removed. The attestation MUST be cryptographically valid, verifiable by the primary key of the certificate in question. This means that a third-party certificate will only be accepted/distributed by the keystore if: the keystore knows about both the first- and third-parties. the third-party has made the identity assertion the first-party has confirmed that they’re OK with the third-party certification being distributed by any keystore. The “ksok” notification is not strictly necessary for this mitigation, but it is intended to avoid potential accidental confusion with any other use of the Third-Party Confirmation signature packet type. The author does not know of any current use that might collide.

defines “Key Server Preferences” with a “No-modify” bit. That bit has never been respected by any keyserver implementation that the author is aware of. An abuse-resistant keystore following effectively treats that bit as always set, whether it is present in the certificate or not.

Creating such an attestation requires multiple steps by different parties, each of which is blocked by all prior steps: The first-party creates the certificate, and transfers it to the third party. The third-party certifies it, and transfers their certification back to the first party. The first party attests to the third party’s certification. Finally, the first party then transfers the compound certificate to the keystore. The complexity and length of such a sequence may represent a usability obstacle to a user who needs a third-party-certified OpenPGP certificate. No current OpenPGP client can easily create the attestations described in this section. More implementation work needs to be done to make it easy (and understandable) for a user to perform this kind of attestation.

An OpenPGP client that needs to interact with an abuse-resistant keystore can take steps to minimize the extent that its interactions with a keystore can be abused by other parties via the attacks described in . This section describes steps that an abuse-resistant client can take.

When performing certificate lookup, an abuse-resistant client SHOULD prefer to query constrained keystores to avoid the risks described in .

When performing lookup by e-mail address, an abuse-resistant client SHOULD consider canonicalizing the e-mail address before searching (see ). When searching by full User ID, unless there is a strong reason to believe that a specific non-normalized form is preferable, an abuse-resistant client SHOULD normalize the entire user ID into Form C (NFC) before performing certificate lookup.

Certificate lookup by arbitrary substring matching, or regular expression is prone to abuse. An abuse-resistant client SHOULD prefer exact-userid or exact-email match lookups where possible. In particular, an abuse-resistant client should avoid trying to offer reliable functionality that performs these sort of fuzzy lookups, and SHOULD warn the user about risks of abuse if the user triggers a codepath that unavoidably performs such a fuzzy lookup.

Key IDs are inherently weaker and easier to spoof or collide than full fingerprints. Where possible, an abuse-resistant keystore client SHOULD use the full fingerprint when interacting with the keystore.

When an abuse-resistant client relies on a keystore for certificate validation, many things can go subtly wrong if the client fails to closely track the specific semantics of the keystore’s validation claims. For example, a certificate published by WKD () at https://openpgpkey.example.org/.well-known/openpgpkey/hu/iy9q119eutrkn8s1mk4r39qejnbu3n5q?l=joe.doe offers a validation claim only for the e-mail address joe.doe@example.org. If the certificate retrieved at that address contains other user IDs, or if the user ID containing that e-mail address contains an display-name, none of that information should be considered “validated” by the fact that the certificate was retrieved via certificate lookup by WKD. When certificate validation is represented more generally by a keystore via certificate retrieval (e.g. from an e-mail validating keyserver that has no distinct certificate validation interface), the thing validated is the certificate received from the keystore, and not the result of the merge into any local copy of the certificate already possessed by the client. Consider also timing and duration of these assertions of validity, which represent a subtle tradeoff between privacy and risk as described in .

An OpenPGP implementation that generates or manages certificates and expects to distribute them via abuse-resistant keystores can take steps to ensure that the certificates generated are more likely to be accessible when needed. This section describes steps such an abuse-sensitive implementation can take.

E-mail addresses can be written in many different ways. An abuse-sensitive implementation considering attaching a user ID containing an e-mail address on a certificate SHOULD ensure that the e-mail address is structured as simply as possible. See for details about e-mail address canonicalization. For example, if the e-mail domain considers the local part to be case-insensitive (as most e-mail domains do today), if a proposed user ID contains the addr-spec: Alice@EXAMPLE.org, the implementation SHOULD warn the user and, if possible, propose replacing the addr-spec part of the user ID with alice@example.org.

User IDs are arbitrary UTF-8 strings, but UTF-8 offers several ways to represent the same string. An abuse-sensitive implementation considering attaching a user ID to a certificate SHOULD normalize the string using Form C (NFC) before creating the self-sig. At the same time, the implementation MAY also warn the user if the “compatibility” normalized form (NFKC) differs from the candidate user ID and, if appropriate, offer to convert the user ID to compatibility normalized form at the user’s discretion.

An abuse-sensitive implementation SHOULD warn the user when attaching a user attribute larger than 8383 octets, and SHOULD refuse to attach user attributes entirely larger than 65536 octets. (See )

requires cross-sigs for all signing-capable subkeys, but is agnostic about the use of cross-sigs for subkeys of other capabilities. An abuse-sensitive implementation that wants a certificate to be discoverable by subkey SHOULD provide cross-sigs for any subkey capable of making a cross-sig.

Issuer subpackets contain only 64-bit key IDs. Issuer Fingerprint subpackets contain an unambiguous designator of the issuing key, avoiding the ambiguities described in . Abuse-sensitive implementations SHOULD providue Issuer Fingerprint subpackets.

Unhashed subpackets may be stripped or mangled. Placing cross-sigs and issuer fingerprint subpackets in the hashed subpackets will ensure that they are propagated by any cryptographically-validating keystore, even if that keystore fails to observe the exceptions in .

If an abuse-sensitive implementation wants other peers to be able to to retrieve the managed certificate by certificate lookup (that is, by searching based on user ID or e-mail address), it needs to be aware that submission to an unrestricted keystore is not reliable (see for more details). Consequently, such an implementation SHOULD submit the managed certificate to restricted, lookup-capable keystores where possible, as those keystores are more likely to be able to offer reliable lookup.

A first-party-only keystore as described in might decline to distribute revocations made by the designated revoker. This is a risk to certificate-holder who depend on this mechanism, because an important revocation might be missed by clients depending on the keystore. FIXME: adjust this document to point out where revocations from a designated revoker SHOULD be propagated, maybe even in first-party-only keystores.

During signature verification, a user performing certificate discovery against a keystore SHOULD prefer to use the full fingerprint as an index, rather than the 64-bit key ID. Using a 64-bit key ID is more likely to run into collision attacks; and if the retrieved certificate has a matching key ID but the signature cannot be validated with it, the client is in an ambiguous state – did it retrieve the wrong certificate, or is the signature incorrect? Using the fingerprint resolves the ambiguity: the signature is incorrect, because the a fingerprint match is overwhelmingly stronger than a key ID match. Unfortunately, many OpenPGP implementations distribute signatures with only an Issuer subpacket, so a client attempting to find such a certificate MAY perform certificate discovery based on only the key ID. A keystore that offers certificate discovery MAY choose to require full fingerprint, but such a keystore will not be useful for a client attempting to verify a bare signature from an unknown party if that signature only has an Issuer subpacket (and no Issuer Fingerprint subpacket).

There are contexts where it is expected and acceptable that the signature appears without its certificate: for example, if the set of valid signers is already known (as in some OpenPGP-signed operating system updates), shipping the certificate alongside the signature would be pointless bloat. However, OpenPGP signatures are often found in contexts where the certificate is not yet known by the verifier. For example, many OpenPGP-signed e-mails are not accompanied by the signing certificate. In another example, the use of authentication-capable OpenPGP keys in standard SSH connections do not contain the full OpenPGP certificates, which means that the SSH clients and servers need to resort to out-of-band processes if evaluation of the OpenPGP certificates is necessary. The certificate discovery interface offered by keystores is an attempt to accommodate these situations. But in the event that a keystore is unavailable, does not know the certificate, or suffers from a flooding attack, signature validation may fail unnecessarily. In an encrypted e-mail context specifically, such a failure may also limit the client’s ability to reply with an encrypted e-mail. Certificate discovery also presents a potential privacy concern for the signature verifier, as noted in . These problematic situations can be mitigated by shipping the certificate in-band, alongside the signature. Signers SHOULD adopt this practice where possible to reduce the dependence of the verifier on the keystores for certificate discovery. is an example of e-mail recommendations that include in-band certificates.

OpenPGP certificates are potenitally large. When distributing an in-band certificate alongside a signature in a context where size is a concern (e.g. bandwidth, latency, or storage costs are a factor), the distributor SHOULD reduce the size of the in-band certificate by stripping unnecessary packets. For example, the distributor may: Strip certification and signature packets that (due to creation and expiration time) are not relevant to the time of the signature itself. This ensures that the reduced certificate is contemporaneously valid with the signature. Strip irrelevant subkeys (and associated Subkey Binding Signature packets and cross-sigs). If the signature was issued by a signing-capable subkey, that subkey (and its binding signature and cross-sig) are clearly relevant. Other signing-capable subkeys are likely to be irrelevant. But determining which other subkeys are relevant may be context-specific. For example, in the e-mail context, an encryption-capable subkey is likely to be contextually relevant, because it enables the recipient to reply encrypted, and therefore should not be stripped. Strip user IDs (and associated certifications) that are unlikely to be relevant to the signature in question. For example, in the e-mail context, strip any user IDs that do not match the declared sender of the message. Strip third-party certifications that are unlikely to be relevant to the verifier. Doing this successfully requires some knowledge about what the third-parties the recipient is likely to care about. Stripping all third-party certifications is a simple means of certificate reduction. The verifier of such a certificate may need to do certificate update against their preferred keystore to learn about third-party certifications useful to them.

Much of this discussion assumes that primary keys are the only certification-capable keys in the OpenPGP ecosystem. Some proposals have been put forward that assume that subkeys can be marked as certification-capable. If subkeys are certification-capable, then much of the reasoning in this draft becomes much more complex, as subkeys themselves can be revoked by their primary key without invalidating the key material itself. That is, a subkey can be both valid (in one context) and invalid (in another context) at the same time. So questions about what data can be dropped (e.g. in ) are much fuzzier, and the underlying assumptions may need to be reviewed. If some OpenPGP implementations accept certification-capable subkeys, but an abuse-resistant keystore does not accept certifications from subkeys in general, then interactions between that keystore and those implementations may be surprising.

Online protocols like TLS perform signature and certificate evaluation based entirely on the present time. If a certificate that signs a TLS handshake message is invalid now, it doesn’t matter whether it was valid a week ago, because the present TLS session is the context of the evaluation. But OpenPGP signatures are often evaluated at some temporal remove from when the signature was made. For example, software packages are signed at release time, but those signatures are validated at download time. A verifier that does not already know the certificate that made the signature will need to perform certificate discovery against some set of keystores to find a certificate with which to check the signature. Further complicating matters, the composable nature of an OpenPGP certificate means that the certificate associated with any particular signing key (primary key or subkey) can transform over time. So when evaluating a signature that appears to have been made by a given certificate, it may be better to try to evaluate the certificate at the time the signature was made, rather than the present time.

When evaluating a certificate at a time T in the past (for example, when trying to validate a data signature by that certificate that was created at time T), one approach is to discard all packets from the certificate if the packet has a creation time later than T. Then evaluate the resulting certificate from the remaining packets in the context of time T. However, any such evaluation MUST NOT ignore “hard” OpenPGP key revocations, regardless of their creation date. (see ).

If a non-append-only keystore () has dropped superseded () or expired () certifications, it’s possible for the certificate composed of the remaining packets to have no valid first-party certification at the time that a given signature was made. If a user performs certificate discovery against such a keystore, the certificate it retrieves would be invalid according to , and consequently verification of any signature by that certificate would fail. One simple mitigation to this problem is to ship a contemporaneously-valid certificate in-band alongside the signature (see ). If the distributor does this, then the verifier need only learn about revocations. If knowledge about revocation is needed, the verifier might perform a certificate update (not “certificate discovery”) against any preferred keystore, including non-append-only keystores, merging what it learns into the in-band contemporary certificate. Then the signature verifier can follow the certificate evaluation process outlined in , using the merged certificate.

The append-only aspect of some OpenPGP keystores encourages a user of the keystore to rely on that keystore as a faithful reporter of history, and one that will not misrepresent or hide the history that they know about. An unfaithful “append-only” keystore could abuse the trust in a number of ways, including withholding revocation certificates, offering different sets of certificates to different clients doing certificate lookup, and so on. However, the most widely used append-only OpenPGP keystore, the keyserver pool, offers no cryptographically verifiable guarantees that it will actually remain append-only. Users of the pool have traditionally relied on its distributed nature, and the presumption that coordination across a wide range of administrators would make it difficult for the pool to reliably lie or omit data. However, the endpoint most commonly used by clients to access the network is hkps://hkps.pool.sks-keyservers.net, the default for . That endpoint is increasingly consolidated, and currently consists of hosts operated by only two distinct administrators, increasing the risk of potential misuse. Offering cryptographic assurances that a keystore could remain append-only is an appealing prospect to defend against these kinds of attack. Many popular schemes for providing such assurances are known as “blockchain” technologies, or global append-only ledgers. With X.509 certificates, we have a semi-functional Certificate Transparency (, or “CT”) ecosystem that is intended to document and preserve evidence of (mis)issuance by well-known certificate authorities (CAs), which implements a type of global append-only ledger. While the CT infrastructure remains vulnerable to certain combinations of colluding actors, it has helped to identify and sanction some failing CAs. Like other global append-only ledgers, CT itself is primarily a detection mechanism, and has no enforcement regime. If a widely-used CA were identified by certificate transparency to be untrustworthy, the rest of the ecosystem still needs to figure out how to impose sanctions or apply a remedy, which may or may not be possible. CT also has privacy implications – the certificates published in the CT logs are visible to everyone, for the lifetime of the log. For spam abatement, CT logs decline all X.509 certificates except those issued by certain CAs (those in popular browser “root stores”). This is an example of the strategy outlined in ). Additional projects that provide some aspects of global append-only ledgers that try to address some of the concerns described here include and , though they are not specific to OpenPGP. Both of these systems are dependent on servers operated by identity providers, however. And both offer the ability to detect a misbehaving identity provider, but no specific enforcement or recovery strategies against such an actor. It’s conceivable that a keystore could piggyback on the CT logs or other blockchain/ledger mechanisms like to store irrevocable pieces of data (such as revocation certificates). Further work is needed to describe how to do this in an effective and performant way.

A typical use case for certificate lookup is a user looking for a certificate in order to be able to encrypt an outbound message intended for a given e-mail address, but this is not the only use case. Another use caes is when the party in control of a particular identity wants to determine whether anyone else is claiming that identity. That is, a client in control of the secret key material associated with a particular certificate with user ID X might search a keystore for all certificates matching X in order to find out whether any other certificates claim it. This is an important safeguard as part of the ledger-based detection mechanisms described in , but may also be useful for keystores in general. However, identity monitoring against a keystore that does not defend against user ID flooding () is expensive and potentially of limited value. In particular, a malicious actor with a certificate which duplicates a given User ID could flood the keystore with similar certificates, hiding whichever one is in malicious use. Since such a keystore is not considered authoritative by any reasonable client for the user ID in question, this attack forces the identity-monitoring defender to spend arbitrary resources fetching and evaluating each certificate in the flood, without knowing which certificate other clients might be evaluating.

This section collects details about common OpenPGP implementation behavior that are useful in evaluating and reasoning about OpenPGP certificates.

It’s useful to classify OpenPGP revocations of key material into two categories: “soft” and “hard”. If the “Reason for Revocation” of an OpenPGP key is either “Key is superseded” or “Key is retired and no longer used”, it is a “soft” revocation. An implementation that interprets a “soft” revocation will typically not invalidate signatures made by the associated key with a creation date that predates the date of the soft revocation. A “soft” revocation in some ways behaves like a non-overridable expiration date. All other revocations of OpenPGP keys (with any other Reason for Revocation, or with no Reason for Revocation at all) should be considered “hard”. The presence of a “hard” revocation of an OpenPGP key indicates that the user should reject all signatures and certifications made by that key, regardless of the creation date of the signature. Note that some OpenPGP implementations do not distinguish between these two categories. A defensive OpenPGP implementation that does not distinguish between these two categories SHOULD treat all revocations as “hard”. An implementation aware of a “soft” revocation or of key or certificate expiry at time T SHOULD accept and process a “hard” revocation even if it appears to have been issued at a time later than T.

requires a user ID to be a UTF-8 string, but does not constrain it beyond that. In practice, a handful of conventions predominate in how User IDs are formed. The most widespread convention is a name-addr as defined in . For example:

]]> But a growing number of OpenPGP certificates contain user IDs that are instead a raw addr-spec, omitting the display-name and the angle brackets entirely, like so:

Some certificates have user IDs that are simply normal human names (perhaps display-name in jargon, though not necessarily conforming to a specific ABNF). For example:

Still other certificates identify a particular network service by scheme and hostname. For example, the administrator of an ssh host participating in the might choose a user ID for the OpenPGP representing the host like so:

When an OpenPGP user IDs includes an addr-spec, there still may be multiple ways of representing the addr-spec that refer to the same underlying mailbox. Having a truly canonical form of an addr-spec is probably impossible because of legacy deployments of mailservers that do odd things with the local part, but e-mail addresses used in an abuse-resistant ecosystem SHOULD be constrained enough to admit to some sensible form of canonicalization.

In , the local-part can be any dot-atom. But if this is decoded, it could be any arbitrary charset, not necessarily UTF-8. FIXME: Do we convert from the arbitrary charset to UTF-8?

FIXME: should domain name be canonicalized into punycode form? User IDs are typically user-facing, so i think the canonicalized form should be the Form C (NFC) of the domain name. Can we punt to some other draft here?

FIXME: recommends downcasing all ASCII characters in the left-hand side, but leaves all

This document offers guidance on mitigating a range of denial-of-service attacks on public keystores, so the entire document is in effect about security considerations. Many of the mitigations described here defend individual OpenPGP certificates against flooding attacks (see ). But only some of these mitigations defend against flooding attacks against the keystore itself (see ), or against flooding attacks on the space of possible user IDs (see ). Thoughtful threat modeling and monitoring of the keystore and its defenses are probably necessary to maintain the long-term health of the keystore. describes a potentially scary security problem for designated revokers. TODO (more security considerations)

Note that there is an inherent tension between accepting arbitrary certificate uploads and permitting effective certificate lookup. If a keystore accepts arbitrary certificate uploads for redistribution, it appears to be vulnerable to user ID flooding (), which makes it difficult or impossible to rely on for certificate lookup. In the broader ecosystem, it may be necessary to use gated/controlled certificate lookup mechanisms. For example, both and enable the administrator of a DNS domain to distribute certificates associated with e-mail addresses within that domain, while excluding other parties. As a rather different example, offers certificate lookup on the basis of interest – a client interested in an organization can use that mechanism to learn what certificates that organization thinks are worth knowing about, associated with a range of identities regardless of the particular DNS domain. Note that does not provide the certificates directly, but instead expects the client to be able to retrieve them by primary key fingerprint through some other keystore capable of (and responsible for) certificate update.

Keystores themselves raise a host of potential privacy concerns. Additional privacy concerns are raised by traffic to and from the keystores. This section tries to outline some of the risks to the privacy of people whose certificates are stored and redistributed in public keystores, as well as risks to the privacy of people who make use of the key stores for certificate lookup, certificate discovery, or certificate update.

Public OpenPGP keystores often distribute names or e-mail addresses of people. Some people do not want their names or e-mail addresses distributed in a public keystore, or may change their minds about it at some point. Append-only keystores are particularly problematic in that regard. The mitigation in can help such users strip their details from keys that they control. However, if an OpenPGP certificate with their details is uploaded to a keystore, but is not under their control, it’s unclear what mechanisms can be used to remove the certificate that couldn’t also be exploited to take down an otherwise valid certificate. Some jurisdictions may present additional legal risk for keystore operators that distribute names or e-mail addresses of non-consenting parties. Updates-only keystores () and user ID redacting keystores () may reduce this particular privacy concern because they distribute no user IDs at all.

Third-party certifications effectively map out some sort of social graph. A certification asserts a statement of belief by the issuer that the real-world party identified by the user ID is in control of the subject cryptographic key material. But those connections may be potentially sensitive, and some people may not want these maps built. A first-party-only keyserver () avoids this privacy concern because it distributes no third-party privacy concern. First-party attested third-party certifications described in are even more relevant edges in the social graph, because their bidirectional nature suggests that both parties are aware of each other, and see some value in mutual association.

Even without third-party certifications, the acts of certificate lookup, certificate discovery, and certificate update represent a potential privacy risk, because the keystore queried gets to learn which user IDs (in the case of lookup) or which certificates (in the case of update or discovery) the client is interested in. In the case of certificate update, if a client attempts to update all of its known certificates from the same keystore, that set is likely to be a unique set, and therefore identifies the client. A keystore that monitors the set of queries it receives might be able to profile or track those clients who use it repeatedly. A privacy-aware client which wants to to avoid such a tracking attack MAY try to perform certificate update from multiple different keystores. To hide network location, a client making a network query to a keystore SHOULD use an anonymity network like . Tools like are designed to facilitate this type of certificate update. Such a client SHOULD also decline to use protocol features that permit linkability across interactions with the same keystore, such as TLS session resumption, HTTP cookies, and so on. Keystores which permit public access and want to protect the privacy of their clients SHOULD NOT reject access from clients using or comparable anonymity networks. Additionally, they SHOULD minimize access logs they retain. Alternately, some keystores may distribute their entire contents to any interested client, in what can be seen as the most trivial form of private information retrieval. is one such example; its contents are distributed as an operating system package. Clients can interrogate their local copy of such a keystore without exposing their queries to a third-party.

If a client relies on a keystore’s certificate validation interface, or on the presence of a certificate in a keystore as a part of its validation calculations, it’s unclear how long the assertion from the keystore is or should be considered to hold. One seemingly simple approach is to simply query the keystore’s validation interface each time that the client needs to validate the certificate. This “live” validation approach poses a quandary to the client in the event that the keystore is unavailable. How should in interpret the “unknown” result? In addition, live validation reveals the client’s activity to the keystore with fine precision. A privacy-aware client that depends on keystores for certificate validation SHOULD NOT perform “live” certificate validation on each use it makes of the certificate. Rather, it SHOULD cache the validation information for some period of time and make use of the cached copy where possible. If such a client does a regular certificate update from the same keystore, it SHOULD also pre-emptively query the keystore for certificate validation at the same time. Choosing the appropriate time intervals for this kind of caching has implications for the windows of risk for the client that might use a revoked certificate. Defining an appropriate schedule to make this tradeoff is beyond the scope of this document.

The act of doing certificate discovery on unknown signatures offers a colluding keystore and remote peer a chance to track a client’s consumption of a given signature. An attacker with access to keystore logs could sign a message with a unique key, and then watch the keystore activity to determine when a client consumes the signature. This is potentially a tracking or “phone-home” situation. A signer that has no interest in this particular form of tracking can mitigate this concern by shipping their certificate in-band, alongside the signature, as recommended in . A privacy-aware client MAY insist on in-band certificates by declining to use any certificate discovery interface at all, and treat a bare signature by an unknown certificate as an invalid signature.

The act of doing certificate update itself reveals some information that the client is interested in a given certificate and how it may have changed since the last time the client updated it, or since it was first received by the client. This is essentially the same privacy problem presented by OCSP in the X.509 world. In the online world of TLS, this privacy leak is mitigated by the CertificateStatus TLS handshake extension (), a.k.a. “OCSP stapling”. There is no comparable solution for the store-and-forward or non-online scenarios where OpenPGP is often found. Privacy-aware clients MAY prefer to access update interfaces from anonymity-preserving networks like to obscure where they are on the network, but if the certificate being updated is known to be used only by a single client that may not help. Privacy-aware clients MAY prefer to stage their certificate updates over time, but longer delays imply greater windows of vulnerability for use of an already-revoked certificate. This strategy also does not help when a previously-unknown certificate is encountered in-band (see ), and the OpenPGP client wants to evaluate it for use in the immediate context.

The distinct keystore interfaces documented in offer subtly different semantics, and are used by a reasonable keystore client at different times. A keystore that offers distinct discovery and update interfaces may infer that a client visiting the update interface already knows about the certificate in question, or that a client visiting the discovery interface is in the process of verifying a signature from a certificate it has not seen before. HKP itself () conflates these two interfaces – the same HKP query is be used to perform both discovery and update (though implementations like are not at all abuse-resistant for either use), which may obscure the context and intent of the client from the keystore somewhat. A privacy-aware client that can afford the additional bandwidth and complexity MAY use the keystore’s discovery interface for both update and discovery, since the discovery interface is a proper superset of the update interface.

If access to the keystore happens over observable channels (e.g., cleartext connections over the Internet), then a passive network monitor could perform the same type profiling or tracking attack against clients of the keystore described in . Keystores which offer network access SHOULD provide encrypted transport.

Even if a keystore offers encrypted transport, the size of queries and responses may provide effective identification of the specific certificates fetched during lookup, discovery, or update, leaving open the types of tracking attacks described in . Clients of keystores SHOULD pad their queries to increase the size of the anonymity set. And keystores SHOULD pad their responses. The appropriate size of padding to effectively anonymize traffic to and from keystores is likely to be mechanism- and cohort-specific. For example, padding for keystores accessed via the DNS ( may use different padding strategies that padding for keystores accessed over WKD (), which may in turn be different from keystores accessed over HKPS (). A keystore which only accepts user IDs within a specific domain (e.g., ) or which uses custom process () for verification might have different padding criteria than a keystore that serves the general public. Specific padding policies or mechanisms are out of scope for this document.

describes some outstanding work that needs to be done to help users understand how to produce and distribute a third-party-certified OpenPGP certificate to an abuse-resistant keystore. Additionally, some keystores present directly user-facing affordances. For example, keyservers typically offer forms and listings that can be viewed directly in a web browser. Such a keystore SHOULD be as clear as possible about what abuse mitigations it takes (or does not take), to avoid user confusion. Keystores which do not expect to be used directly as part of a certificate validation calculation SHOULD advise clients as explicitly as possible that they offer no assertions of validity. Experience with the keyserver network shows that many users treat the keyserver web interfaces as authoritative. That is, users act as though the keyserver network offers some type of certificate validation. Unfortunately, The developer and implementor communities explicitly disavow any authoritative role in the ecosystem, and the implementations attempt very few mitigations against abuse, permitting redistribution of even cryptographically invalid OpenPGP packets. Clearer warnings to end users might reduce this kind of misperception. Or the community could encourage the removal of frequently misinterpreted user interfaces entirely.

This document asks IANA to register two entries in the OpenPGP Notation IETF namespace, both with a reference to this document: the “ksok” notation is defined in . the “uidhash” notation is defined in .

[ RFC Editor: please remove this section before publication ] This document is currently edited as markdown. Minor editorial changes can be suggested via merge requests at https://gitlab.com/dkg/draft-openpgp-abuse-resistant-keystore or by e-mail to the author. Please direct all significant commentary to the public IETF OpenPGP mailing list: openpgp@ietf.org

substantive changes between -02 and -03: new sections: Keystore Interfaces Keystore Client Best Practices Certificate Generation and Management Best Practices rename “certificate discovery” to “certificate lookup” redefine “certificate discovery” to refer to lookup by signing (sub)key new attack: fingerprint flooding new retrieval-time mitigations – tighter filters on discovery and update recommend in-band certificates where possible to avoid discovery and lookup new privacy considerations: distinct keystore interfaces certificate update certificate discovery certificate validation more nuance about unhashed subpacket filtering substantive changes between -01 and -02: distinguish different forms of flooding attack distinguish toxic data as distinct from flooding retrieval-time mitigations user ID redaction references to related work (CT, keylist, CONIKS, key transparency, ledgers/”blockchain”, etc) more details about UI/UX substantive changes between -00 and -01: split out Contextual and Non-Append-Only mitigations documented several other mitigations, including: Decline Data From the Future Blocklist Exterior Process Designated Authorities Known Certificates Rate-Limiting Scoped User IDs documented Updates-Only Keystores consider three different kinds of flooding deeper discussion of privacy considerations better documentation of Reason for Revocation document user ID conventions

This document is the result of years of operational experience and observation, as well as conversations with many different people – users, implementors, keystore operators, etc. A non-exhaustive list of people who have contributed ideas or nuance to this document specifically includes: Antoine Beaupré ilf Jamie McClelland Jonathan McDowell Justus Winter Marcus Brinkmann Micah Lee Neal Walfield Phil Pennock Philihp Busby vedaal Vincent Breitmoser Wiktor Kwapisiewicz