Google

345 Spear Street San Francisco, CA 94105 US bittau@google.com

Stanford University

353 Serra Mall, Room 475 Stanford, CA 94305 US dabo@cs.stanford.edu

Stanford University

353 Serra Mall, Room 288 Stanford, CA 94305 US dbg@scs.stanford.edu

Stanford University

353 Serra Mall, Room 475 Stanford, CA 94305 US mike@shiftleft.org

University College London

Gower St. London WC1E 6BT UK M.Handley@cs.ucl.ac.uk

Stanford University

353 Serra Mall, Room 290 Stanford, CA 94305 US dm@uun.org

Stanford University

353 Serra Mall, Room 288 Stanford, CA 94305 US sqs@cs.stanford.edu

Kestrel Institute

3260 Hillview Avenue Palo Alto, CA 94304 US eric.smith@kestrel.edu

Internet tcp encryption This document specifies tcpcrypt, a TCP encryption protocol designed for use in conjunction with the TCP Encryption Negotiation Option (TCP-ENO) . Tcpcrypt coexists with middleboxes by tolerating resegmentation, NATs, and other manipulations of the TCP header. The protocol is self-contained and specifically tailored to TCP implementations, which often reside in kernels or other environments in which large external software dependencies can be undesirable. Because the size of TCP options is limited, the protocol requires one additional one-way message latency to perform key exchange before application data may be transmitted. However, this cost can be avoided between two hosts that have recently established a previous tcpcrypt connection.

The key words "MUST", "MUST NOT", "REQUIRED", "SHALL", "SHALL NOT", "SHOULD", "SHOULD NOT", "RECOMMENDED", "MAY", and "OPTIONAL" in this document are to be interpreted as described in .

This document describes tcpcrypt, an extension to TCP for cryptographic protection of session data. Tcpcrypt was designed to meet the following goals: Meet the requirements of the TCP Encryption Negotiation Option (TCP-ENO) for protecting connection data. Be amenable to small, self-contained implementations inside TCP stacks. Minimize additional latency at connection startup. As much as possible, prevent connection failure in the presence of NATs and other middleboxes that might normalize traffic or otherwise manipulate TCP segments. Operate independently of IP addresses, making it possible to authenticate resumed sessions efficiently even when either end changes IP address.

This section describes the tcpcrypt protocol at an abstract level. The concrete format of all messages is specified in .

Setting up a tcpcrypt connection employs three types of cryptographic algorithms: A key agreement scheme is used with a short-lived public key to agree upon a shared secret. An extract function is used to generate a pseudo-random key from some initial keying material, typically the output of the key agreement scheme. The notation Extract(S, IKM) denotes the output of the extract function with salt S and initial keying material IKM. A collision-resistant pseudo-random function (CPRF) is used to generate multiple cryptographic keys from a pseudo-random key, typically the output of the extract function. We use the notation CPRF(K, CONST, L) to designate the output of L bytes of the pseudo-random function identified by key K on CONST. The Extract and CPRF functions used by default are the Extract and Expand functions of HKDF . These are defined as follows in terms of the PRF HMAC-Hash(key, value) for a negotiated Hash function:

HKDF-Extract(salt, IKM) -> PRK PRK = HMAC-Hash(salt, IKM) HKDF-Expand(PRK, CONST, L) -> OKM T(0) = empty string (zero length) T(1) = HMAC-Hash(PRK, T(0) | CONST | 0x01) T(2) = HMAC-Hash(PRK, T(1) | CONST | 0x02) T(3) = HMAC-Hash(PRK, T(2) | CONST | 0x03) ... OKM = first L octets of T(1) | T(2) | T(3) | ...

Lastly, once tcpcrypt has been successfully set up and encryption keys have been derived, an algorithm for Authenticated Encryption with Associated Data (AEAD) is used to protect the confidentiality and integrity of all transmitted application data. AEAD algorithms use a single key to encrypt their input data and also to generate a cryptographic tag to accompany the resulting ciphertext; when decryption is performed, the tag allows authentication of the encrypted data and of optional, associated plaintext data.

Tcpcrypt depends on TCP-ENO to negotiate whether encryption will be enabled for a connection, and also which key agreement scheme to use. TCP-ENO negotiates the use of a particular TCP encryption protocol or TEP by including protocol identifiers in ENO suboptions. This document associates four TEP identifiers with the tcpcrypt protocol, as listed in . Each identifier indicates the use of a particular key-agreement scheme. Future standards may associate additional identifiers with tcpcrypt. An active opener that wishes to negotiate the use of tcpcrypt includes an ENO option in its SYN segment. That option includes suboptions with tcpcrypt TEP identifiers indicating the key-agreement schemes it is willing to enable. The active opener MAY additionally include suboptions indicating support for encryption protocols other than tcpcrypt, as well as global suboptions as specified by TCP-ENO. If a passive opener receives an ENO option including tcpcrypt TEPs it supports, it MAY then attach an ENO option to its SYN-ACK segment, including solely the TEP it wishes to enable. To establish distinct roles for the two hosts in each connection, tcpcrypt depends on the role-negotiation mechanism of TCP-ENO. As one result of the negotiation process, TCP-ENO assigns hosts unique roles abstractly called "A" at one end of the connection and "B" at the other. Generally, an active opener plays the "A" role and a passive opener plays the "B" role; but in the case of simultaneous open, an additional mechanism breaks the symmetry and assigns different roles to the two hosts. This document adopts the terms "host A" and "host B" to identify each end of a connection uniquely, following TCP-ENO's designation. ENO suboptions include a flag v which indicates the presence of associated, variable-length data. In order to propose fresh key agreement with a particular tcpcrypt TEP, a host sends a one-byte suboption containing the TEP identifier and v = 0. In order to propose session resumption (described further below) with a particular TEP, a host sends a variable-length suboption containing the TEP identifier, the flag v = 1, and an identifier for a session previously negotiated with the same host and the same TEP. Once two hosts have exchanged SYN segments, TCP-ENO defines the negotiated TEP to be the last valid TEP identifier in the SYN segment of host B (that is, the passive opener in the absence of simultaneous open) that also occurs in that of host A. If there is no such TEP, hosts MUST disable TCP-ENO and tcpcrypt. If the negotiated TEP was sent by host B with v = 0, it means that fresh key agreement will be performed as described below in . If it had v = 1, the key-exchange messages will be omitted in favor of determining keys via session-caching as described in , and protected application data may immediately be sent as detailed in . Note that the negotiated TEP is determined without reference to the v bits in ENO suboptions, so if host A offers resumption with a particular TEP and host B replies with a non-resumption suboption with the same TEP, that may become the negotiated TEP and fresh key agreement will be performed. That is, sending a resumption suboption also implies willingness to perform fresh key agreement with the indicated TEP. As required by TCP-ENO, once a host has both sent and received an ACK segment containing a valid ENO option, encryption MUST be enabled and plaintext application data MUST NOT ever be exchanged on the connection. If the negotiated TEP is among those listed in , a host MUST follow the protocol described in this document.

Following successful negotiation of a tcpcrypt TEP, all further signaling is performed in the Data portion of TCP segments. Except when resumption was negotiated (described below in ), the two hosts perform key exchange through two messages, Init1 and Init2, at the start of the data streams of host A and host B, respectively. These messages may span multiple TCP segments and need not end at a segment boundary. However, the segment containing the last byte of an Init1 or Init2 message SHOULD have TCP's PSH bit set. The key exchange protocol, in abstract, proceeds as follows:

A -> B: Init1 = { INIT1_MAGIC, sym-cipher-list, N_A, PK_A } B -> A: Init2 = { INIT2_MAGIC, sym-cipher, N_B, PK_B }

The concrete format of these messages is specified in . The parameters are defined as follows: INIT1_MAGIC, INIT2_MAGIC: constants defined in . sym-cipher-list: a list of symmetric ciphers (AEAD algorithms) acceptable to host A. These are specified in . sym-cipher: the symmetric cipher selected by host B from the sym-cipher-list sent by host A. N_A, N_B: nonces chosen at random by hosts A and B, respectively. PK_A, PK_B: ephemeral public keys for hosts A and B, respectively. These, as well as their corresponding private keys, are short-lived values that SHOULD be refreshed periodically. The private keys SHOULD NOT ever be written to persistent storage. The ephemeral secret (ES) is the result of the key-agreement algorithm (see ) indicated by the negotiated TEP. The inputs to the algorithm are the local host's ephemeral private key and the remote host's ephemeral public key. For example, host A would compute ES using its own private key (not transmitted) and host B's public key, PK_B. The two sides then compute a pseudo-random key (PRK), from which all session keys are derived, as follows:

PRK = Extract(N_A, eno-transcript | Init1 | Init2 | ES)

Above, | denotes concatenation; eno-transcript is the protocol-negotiation transcript defined in TCP-ENO; and Init1 and Init2 are the transmitted encodings of the messages described in . A series of "session secrets" are then computed from PRK as follows:

ss[0] = PRK ss[i] = CPRF(ss[i-1], CONST_NEXTK, K_LEN)

The value ss[0] is used to generate all key material for the current connection. The values ss[i] for i > 0 can be used to avoid public key cryptography when establishing subsequent connections between the same two hosts, as described in . The CONST_* values are constants defined in . The length K_LEN depends on the tcpcrypt TEP in use, and is specified in . Given a session secret ss, the two sides compute a series of master keys as follows:

mk[0] = CPRF(ss, CONST_REKEY, K_LEN) mk[i] = CPRF(mk[i-1], CONST_REKEY, K_LEN)

The particular master key in use is advanced as described in . Finally, each master key mk is used to generate keys for authenticated encryption for the "A" and "B" roles. Key k_ab is used by host A to encrypt and host B to decrypt, while k_ba is used by host B to encrypt and host A to decrypt.

k_ab = CPRF(mk, CONST_KEY_A, ae_keylen) k_ba = CPRF(mk, CONST_KEY_B, ae_keylen)

The value ae_keylen depends on the authenticated-encryption algorithm selected, and is given under "Key Length" in . After host B sends Init2 or host A receives it, that host may immediately begin transmitting protected application data as described in . If host A receives Init2 with a sym-cipher value that was not present in the sym-cipher-list it previously transmitted in Init1, it MUST abort the connection and raise an error condition distinct from the end-of-file condition.

TCP-ENO requires each TEP to define a session ID value that uniquely identifies each encrypted connection. As required, a tcpcrypt session ID begins with the negotiated TEP identifier along with the v bit as transmitted by host B. The remainder of the ID is derived from the session secret, as follows:

session_id[i] = TEP-byte | CPRF(ss[i], CONST_SESSID, K_LEN)

Again, the length K_LEN depends on the TEP, and is specified in .

When two hosts have already negotiated session secret ss[i-1], they can establish a new connection without public-key operations using ss[i]. A host signals willingness to resume with a particular session secret by sending a SYN segment with a resumption suboption: that is, an ENO suboption containing the negotiated TEP identifier from the original session and part of an identifier for the session. The resumption identifier is calculated from a session secret ss[i] as follows:

resume[i] = CPRF(ss[i], CONST_RESUME, 18)

To name a session for resumption, a host sends either the first or second half of the resumption identifier, according to the role it played in the original session with secret ss[0]. A host that originally played role A and wishes to resume from a cached session sends a suboption with the first half of the resumption identifier:

byte 0 1 9 (10 bytes total) +--------+--------+---...---+--------+ | TEP- | resume[i]{0..8} | | byte | | +--------+--------+---...---+--------+

Similarly, a host that originally played role B sends a suboption with the second half of the resumption identifier:

byte 0 1 9 (10 bytes total) +--------+--------+---...---+--------+ | TEP- | resume[i]{9..17} | | byte | | +--------+--------+---...---+--------+

If a passive opener recognizes the identifier-half in a resumption suboption it has received and knows ss[i], it SHOULD (with exceptions specified below) agree to resume from the cached session by sending its own resumption suboption, which will contain the other half of the identifier. If it does not agree to resumption with a particular TEP, the passive opener may either request fresh key exchange by responding with a non-resumption suboption using the same TEP, or else respond to any other received suboption. If an active opener receives a resumption suboption for a particular TEP and the received identifier-half does not match the resume[i] value whose other half it previously sent in a resumption suboption for the same TEP, it MUST ignore that suboption. In the typical case that this was the only ENO suboption received, this means the host MUST disable TCP-ENO and tcpcrypt: that is, it MUST NOT send any more ENO options and MUST NOT encrypt the connection. When a host concludes that TCP-ENO negotiation has succeeded for some TEP that was received in a resumption suboption, it MUST then enable encryption with that TEP, using the cached session secret, as described in . The session ID () is constructed in the same way for resumed sessions as it is for fresh ones. In this case the first byte will always have v = 1. The remainder of the ID is derived from the cached session secret. In the case of simultaneous open where TCP-ENO is able to establish asymmetric roles, two hosts that simultaneously send SYN segments with compatible resumption suboptions may resume the associated session. In a particular SYN segment, a host SHOULD NOT send more than one resumption suboption, and MUST NOT send more than one resumption suboption with the same TEP identifier. But in addition to any resumption suboptions, an active opener MAY include non-resumption suboptions describing other key-agreement schemes it supports (in addition to that indicated by the TEP in the resumption suboption). After using ss[i] to compute mk[0], implementations SHOULD compute and cache ss[i+1] for possible use by a later session, then erase ss[i] from memory. Hosts SHOULD retain ss[i+1] until it is used or the memory needs to be reclaimed. Hosts SHOULD NOT write a cached ss[i+1] value to non-volatile storage. When proposing resumption, the active opener MUST use the lowest value of i that has not already been used (successfully or not) to negotiate resumption with the same host and for the same pre-session key ss[0]. A host MUST NOT resume with a session secret if it has ever successfully negotiated resumption in the past, in either role, with the same secret. In the event that two hosts simultaneously send SYN segments to each other that propose resumption with the same session secret but the two segments are not part of a simultaneous open, both connections will have to revert to fresh key-exchange. To avoid this limitation, implementations MAY choose to implement session caching such that a given pre-session key ss[0] is only used for either passive or active opens at the same host, not both. When two hosts have previously negotiated a tcpcrypt session, either host may initiate session resumption regardless of which host was the active opener or played the "A" role in the previous session. However, a given host must either encrypt with k_ab for all sessions derived from the same pre-session key ss[0], or with k_ba. Thus, which keys a host uses to send segments is not affected by the role it plays in the current connection: it depends only on whether the host played the "A" or "B" role in the initial session. Implementations that perform session caching MUST provide a means for applications to control session caching, including flushing cached session secrets associated with an ESTABLISHED connection or disabling the use of caching for a particular connection.

Following key exchange (or its omission via session caching), all further communication in a tcpcrypt-enabled connection is carried out within delimited application frames that are encrypted and authenticated using the agreed keys. This protection is provided via algorithms for Authenticated Encryption with Associated Data (AEAD). The particular algorithms that may be used are listed in . One algorithm is selected during the negotiation described in . The format of an application frame is specified in . A sending host breaks its stream of application data into a series of chunks. Each chunk is placed in the data portion of a "plaintext" value, which is then encrypted to yield a frame's ciphertext field. Chunks must be small enough that the ciphertext (whose length depends on the AEAD cipher used, and is generally slightly longer than the plaintext) has length less than 2^16 bytes. An "associated data" value (see ) is constructed for the frame. It contains the frame's control field and the length of the ciphertext. A "frame nonce" value (see ) is also constructed for the frame but not explicitly transmitted. It contains an offset field whose integer value is the zero-indexed byte offset of the beginning of the current application frame in the underlying TCP datastream. (That is, the offset in the framing stream, not the plaintext application stream.) Because it is strictly necessary for the security of the AEAD algorithm, an implementation MUST NOT ever transmit distinct frames with the same nonce value under the same encryption key. In particular, a retransmitted TCP segment MUST contain the same payload bytes for the same TCP sequence numbers, and a host MUST NOT transmit more than 2^64 bytes in the underlying TCP datastream (which would cause the offset field to wrap) before re-keying. With reference to the "AEAD Interface" described in Section 2 of , tcpcrypt invokes the AEAD algorithm with the secret key K set to k_ab or k_ba, according to the host's role as described in . The plaintext value serves as P, the associated data as A, and the frame nonce as N. The output of the encryption operation, C, is transmitted in the frame's ciphertext field. When a frame is received, tcpcrypt reconstructs the associated data and frame nonce values (the former contains only data sent in the clear, and the latter is implicit in the TCP stream), and provides these and the ciphertext value to the the AEAD decryption operation. The output of this operation is either a plaintext value P or the special symbol FAIL. In the latter case, the implementation MUST either ignore the frame or abort the connection; but if it aborts, the implementation MUST raise an error condition distinct from the end-of-file condition.

The ciphertext field of the application frame contains protected versions of certain TCP header values. When URGp is set, the urgent value indicates an offset from the current frame's beginning offset; the sum of these offsets gives the index of the last byte of urgent data in the application datastream. When FINp is set, it indicates that the sender will send no more application data after this frame. When the TCP FIN flag differs from FINp, a receiving host MUST either ignore the segment altogether or abort the connection and raise an error condition distinct from the end-of-file condition.

Re-keying allows hosts to wipe from memory keys that could decrypt previously transmitted segments. It also allows the use of AEAD ciphers that can securely encrypt only a bounded number of messages under a given key. We refer to the two encryption keys (k_ab, k_ba) as a key-set. We refer to the key-set generated by mk[i] as the key-set with generation number i within a session. Each host maintains a local generation number that determines which key-set it uses to encrypt outgoing frames, and a remote generation number equal to the highest generation used in frames received from its peer. Initially, these two values are set to zero. A host MAY increment its local generation number beyond the remote generation number it has recorded. We call this action initiating re-keying. When a host has incremented its local generation number and uses the new key-set for the first time to encrypt an outgoing frame, it MUST set rekey = 1 for that frame. It MUST set this field to zero in all other cases. When a host receives a frame with rekey = 1, it increments its record of the remote generation number. If the remote generation number is now greater than the local generation number, the receiver MUST immediately increment its local generation number to match and send a frame (with empty application data if necessary) with rekey = 1. A host SHOULD NOT initiate more than one concurrent re-key operation if it has no data to send; that is, it should not initiate re-keying with an empty application frame more than once while its record of the remote generation number is less than its own. When retransmitting, implementations must always transmit the same bytes for the same TCP sequence numbers. Thus, a frame in a retransmitted segment MUST always be encrypted with the same key as when it was originally transmitted. Implementations SHOULD delete older-generation keys from memory once they have received all frames they will need to decrypt with the old keys and have encrypted all outgoing frames under the old keys.

Instead of using TCP Keep-Alives to verify that the remote endpoint is still responsive, tcpcrypt implementations SHOULD employ the re-keying mechanism for this purpose, as follows. When necessary, a host SHOULD probe the liveness of its peer by initiating re-keying and transmitting a new frame immediately (with empty application data if necessary). As described in , a host receiving a frame encrypted under a generation number greater than its own MUST increment its own generation number and immediately transmit a new frame (with zero-length application data, if necessary). Implementations MAY use TCP Keep-Alives for purposes that do not require endpoint authentication, as discussed in .

This section provides byte-level encodings for values transmitted or computed by the protocol.

The Init1 message has the following encoding:

byte 0 1 2 3 +-------+-------+-------+-------+ | INIT1_MAGIC | | | +-------+-------+-------+-------+ 4 5 6 7 +-------+-------+-------+-------+ | message_len | | = M | +-------+-------+-------+-------+ 8 +--------+-------+-------+---...---+-------+ |nciphers|sym- |sym- | |sym- | | =K+1 |cipher0|cipher1| |cipherK| +--------+-------+-------+---...---+-------+ K + 10 K + 10 + N_A_LEN | | v v +-------+---...---+-------+-------+---...---+-------+ | N_A | PK_A | | | | +-------+---...---+-------+-------+---...---+-------+ M - 1 +-------+---...---+-------+ | ignored | | | +-------+---...---+-------+

The constant INIT1_MAGIC is defined in . The four-byte field message_len gives the length of the entire Init1 message, encoded as a big-endian integer. The nciphers field contains an integer value that specifies the number of one-byte symmetric-cipher identifiers that follow. The sym-cipher bytes identify cryptographic algorithms in . The length N_A_LEN and the length of PK_A are both determined by the negotiated key-agreement scheme, as described in . When sending Init1, implementations of this protocol MUST omit the field ignored; that is, they must construct the message such that its end, as determined by message_len, coincides with the end of the field PK_A. When receiving Init1, however, implementations MUST permit and ignore any bytes following PK_A. The Init2 message has the following encoding:

byte 0 1 2 3 +-------+-------+-------+-------+ | INIT2_MAGIC | | | +-------+-------+-------+-------+ 4 5 6 7 8 +-------+-------+-------+-------+-------+ | message_len |sym- | | = M |cipher | +-------+-------+-------+-------+-------+ 9 9 + N_B_LEN | | v v +-------+---...---+-------+-------+---...---+-------+ | N_B | PK_B | | | | +-------+---...---+-------+-------+---...---+-------+ M - 1 +-------+---...---+-------+ | ignored | | | +-------+---...---+-------+

The constant INIT2_MAGIC is defined in . The four-byte field message_len gives the length of the entire Init2 message, encoded as a big-endian integer. The sym-cipher value is a selection from the symmetric-cipher identifiers in the previously-received Init1 message. The length N_B_LEN and the length of PK_B are both determined by the negotiated key-agreement scheme, as described in . When sending Init2, implementations of this protocol MUST omit the field ignored; that is, they must construct the message such that its end, as determined by message_len, coincides with the end of the PK_B field. When receiving Init2, however, implementations MUST permit and ignore any bytes following PK_B.

An application frame comprises a control byte and a length-prefixed ciphertext value:

byte 0 1 2 3 clen+2 +-------+-------+-------+-------+---...---+-------+ |control| clen | ciphertext | +-------+-------+-------+-------+---...---+-------+

The field clen is an integer in big-endian format and gives the length of the ciphertext field. The byte control has this structure:

bit 7 1 0 +-------+---...---+-------+-------+ | cres | rekey | +-------+---...---+-------+-------+

The seven-bit field cres is reserved; implementations MUST set these bits to zero when sending, and MUST ignore them when receiving. The use of the rekey field is described in .

The ciphertext field is the result of applying the negotiated authenticated-encryption algorithm to a "plaintext" value, which has one of these two formats:

byte 0 1 plen-1 +-------+-------+---...---+-------+ | flags | data | +-------+-------+---...---+-------+ byte 0 1 2 3 plen-1 +-------+-------+-------+-------+---...---+-------+ | flags | urgent | data | +-------+-------+-------+-------+---...---+-------+

(Note that clen in the previous section will generally be greater than plen, as the ciphertext produced by the authenticated-encryption scheme must both encrypt the application data and provide a way to verify its integrity.) The flags byte has this structure:

bit 7 6 5 4 3 2 1 0 +----+----+----+----+----+----+----+----+ | fres |URGp|FINp| +----+----+----+----+----+----+----+----+

The six-bit value fres is reserved; implementations MUST set these six bits to zero when sending, and MUST ignore them when receiving. When the URGp bit is set, it indicates that the urgent field is present, and thus that the plaintext value has the second structure variant above; otherwise the first variant is used. The meaning of urgent and of the flag bits is described in .

An application frame's "associated data" (which is supplied to the AEAD algorithm when decrypting the ciphertext and verifying the frame's integrity) has this format:

byte 0 1 2 +-------+-------+-------+ |control| clen | +-------+-------+-------+

It contains the same values as the frame's control and clen fields.

Lastly, a "frame nonce" (provided as input to the AEAD algorithm) has this format:

byte +------+------+------+------+ 0 | FRAME_NONCE_MAGIC | +------+------+------+------+ 4 | | + offset + 8 | | +------+------+------+------+

The 4-byte magic constant is defined in . The 8-byte offset field contains an integer in big-endian format. Its value is specified in .

The TEP negotiated via TCP-ENO may indicate the use of one of the key-agreement schemes named in . For example, TCPCRYPT_ECDHE_P256 names the tcpcrypt protocol with key-agreement scheme ECDHE-P256. All schemes listed there use HKDF-Expand-SHA256 as the CPRF, and these lengths for nonces and session keys:

N_A_LEN: 32 bytes N_B_LEN: 32 bytes K_LEN: 32 bytes

Key-agreement schemes ECDHE-P256 and ECDHE-P521 employ the ECSVDP-DH secret value derivation primitive defined in . The named curves are defined in . When the public-key values PK_A and PK_B are transmitted as described in , they are encoded with the "Elliptic Curve Point to Octet String Conversion Primitive" described in Section E.2.3 of , and are prefixed by a two-byte length in big-endian format:

byte 0 1 2 L - 1 +-------+-------+-------+---...---+-------+ | pubkey_len | pubkey | | = L | | +-------+-------+-------+---...---+-------+

Implementations SHOULD encode these pubkey values in "compressed format", and MUST accept values encoded in "compressed", "uncompressed" or "hybrid" formats. Key-agreement schemes ECDHE-Curve25519 and ECDHE-Curve448 use the functions X25519 and X448, respectively, to perform the Diffie-Helman protocol as described in . When using these ciphers, public-key values PK_A and PK_B are transmitted directly with no length prefix: 32 bytes for Curve25519, and 56 bytes for Curve448. A tcpcrypt implementation MUST support at least the schemes ECDHE-P256 and ECDHE-P521, although system administrators need not enable them.

Specifiers and key-lengths for AEAD algorithms are given in . The algorithms AEAD_AES_128_GCM and AEAD_AES_256_GCM are specified in . The algorithm AEAD_CHACHA20_POLY1305 is specified in .

Tcpcrypt's TEP identifiers will need to be incorporated in IANA's TCP-ENO encryption protocol identifier registry, as follows: cs Spec name 0x21TCPCRYPT_ECDHE_P256 0x22TCPCRYPT_ECDHE_P521 0x23TCPCRYPT_ECDHE_Curve25519 0x24TCPCRYPT_ECDHE_Curve448 A "tcpcrypt AEAD parameter" registry needs to be maintained by IANA as in the following table. The use of encryption is described in . AEAD Algorithm Key Length sym-cipher AEAD_AES_128_GCM16 bytes0x01 AEAD_AES_256_GCM32 bytes0x02 AEAD_CHACHA20_POLY130532 bytes0x10

Public-key generation, public-key encryption, and shared-secret generation all require randomness. Other tcpcrypt functions may also require randomness, depending on the algorithms and modes of operation selected. A weak pseudo-random generator at either host will compromise tcpcrypt's security. Many of tcpcrypt's cryptographic functions require random input, and thus any host implementing tcpcrypt MUST have access to a cryptographically-secure source of randomness or pseudo-randomness. Most implementations will rely on system-wide pseudo-random generators seeded from hardware events and a seed carried over from the previous boot. Once a pseudo-random generator has been properly seeded, it can generate effectively arbitrary amounts of pseudo-random data. However, until a pseudo-random generator has been seeded with sufficient entropy, not only will tcpcrypt be insecure, it will reveal information that further weakens the security of the pseudo-random generator, potentially harming other applications. As required by TCP-ENO, implementations MUST NOT send ENO options unless they have access to an adequate source of randomness. The cipher-suites specified in this document all use HMAC-SHA256 to implement the collision-resistant pseudo-random function denoted by CPRF. A collision-resistant function is one on which, for sufficiently large L, an attacker cannot find two distinct inputs K_1, CONST_1 and K_2, CONST_2 such that CPRF(K_1, CONST_1, L) = CPRF(K_2, CONST_2, L). Collision resistance is important to assure the uniqueness of session IDs, which are generated using the CPRF. All of the security considerations of TCP-ENO apply to tcpcrypt. In particular, tcpcrypt does not protect against active eavesdroppers unless applications authenticate the session ID. If it can be established that the session IDs computed at each end of the connection match, then tcpcrypt guarantees that no man-in-the-middle attacks occurred unless the attacker has broken the underlying cryptographic primitives (e.g., ECDH). A proof of this property for an earlier version of the protocol has been published . To gain middlebox compatibility, tcpcrypt does not protect TCP headers. Hence, the protocol is vulnerable to denial-of-service from off-path attackers. Possible attacks include desynchronizing the underlying TCP stream, injecting RST packets, and forging or suppressing rekey bits. These attacks will cause a tcpcrypt connection to hang or fail with an error. Implementations MUST give higher-level software a way to distinguish such errors from a clean end-of-stream (indicated by an authenticated FINp bit) so that applications can avoid semantic truncation attacks. There is no "key confirmation" step in tcpcrypt. This is not required because tcpcrypt's threat model includes the possibility of a connection to an adversary. If key negotiation is compromised and yields two different keys, all subsequent frames will be ignored due to failed integrity checks, causing the application's connection to hang. This is not a new threat because in plain TCP, an active attacker could have modified sequence and acknowledgement numbers to hang the connection anyway. Tcpcrypt uses short-lived public keys to provide forward secrecy. All currently specified key agreement schemes involve ECDHE-based key agreement, meaning a new key can be efficiently computed for each connection. If implementations reuse these parameters, they SHOULD limit the lifetime of the private parameters, ideally to no more than two minutes. Attackers cannot force passive openers to move forward in their session caching chain without guessing the content of the resumption identifier, which will be difficult without key knowledge.

Tcpcrypt transforms a shared pseudo-random key (PRK) into cryptographic session keys for each direction. Doing so requires an asymmetry in the protocol, as the key derivation function must be perturbed differently to generate different keys in each direction. Tcpcrypt includes other asymmetries in the roles of the two hosts, such as the process of negotiating algorithms (e.g., proposing vs. selecting cipher suites).

Many hosts implement TCP Keep-Alives as an option for applications to ensure that the other end of a TCP connection still exists even when there is no data to be sent. A TCP Keep-Alive segment carries a sequence number one prior to the beginning of the send window, and may carry one byte of "garbage" data. Such a segment causes the remote side to send an acknowledgment. Unfortunately, tcpcrypt cannot cryptographically verify Keep-Alive acknowledgments. Hence, an attacker could prolong the existence of a session at one host after the other end of the connection no longer exists. (Such an attack might prevent a process with sensitive data from exiting, giving an attacker more time to compromise a host and extract the sensitive data.) Thus, tcpcrypt specifies a way to stimulate the remote host to send verifiably fresh and authentic data, described in . The TCP keep-alive mechanism has also been used for its effects on intermediate nodes in the network, such as preventing flow state from expiring at NAT boxes or firewalls. As these purposes do not require the authentication of endpoints, implementations may safely accomplish them using either the existing TCP keep-alive mechanism or tcpcrypt's verified keep-alive mechanism.

We are grateful for contributions, help, discussions, and feedback from the TCPINC working group, including Marcelo Bagnulo, David Black, Bob Briscoe, Jana Iyengar, Tero Kivinen, Mirja Kuhlewind, Yoav Nir, Christoph Paasch, Eric Rescorla, and Kyle Rose. This work was funded by gifts from Intel (to Brad Karp) and from Google; by NSF award CNS-0716806 (A Clean-Slate Infrastructure for Information Flow Control); by DARPA CRASH under contract #N66001-10-2-4088; and by the Stanford Secure Internet of Things Project.

Value Name 0x01CONST_NEXTK 0x02CONST_SESSID 0x03CONST_REKEY 0x04CONST_KEY_A 0x05CONST_KEY_B 0x06CONST_RESUME 0x15101a0eINIT1_MAGIC 0x097105e0INIT2_MAGIC 0x44415441FRAME_NONCE_MAGIC