idnits 2.17.1 draft-taddei-smart-cless-introduction-01.txt: Checking boilerplate required by RFC 5378 and the IETF Trust (see https://trustee.ietf.org/license-info): ---------------------------------------------------------------------------- No issues found here. Checking nits according to https://www.ietf.org/id-info/1id-guidelines.txt: ---------------------------------------------------------------------------- No issues found here. Checking nits according to https://www.ietf.org/id-info/checklist : ---------------------------------------------------------------------------- No issues found here. Miscellaneous warnings: ---------------------------------------------------------------------------- == The copyright year in the IETF Trust and authors Copyright Line does not match the current year -- The document date (July 08, 2019) is 1754 days in the past. Is this intentional? Checking references for intended status: Informational ---------------------------------------------------------------------------- == Unused Reference: 'EPTAXONOMY' is defined on line 1641, but no explicit reference was found in the text == Outdated reference: A later version (-02) exists of draft-mcfadden-smart-endpoint-taxonomy-for-cless-00 == Outdated reference: A later version (-02) exists of draft-mcfadden-smart-endpoint-taxonomy-for-cless-00 -- Duplicate reference: draft-mcfadden-smart-endpoint-taxonomy-for-cless, mentioned in 'I-D.draft-mcfadden-smart-endpoint-taxonomy-for-cless-00', was also mentioned in 'EPTAXONOMY'. Summary: 0 errors (**), 0 flaws (~~), 4 warnings (==), 2 comments (--). Run idnits with the --verbose option for more detailed information about the items above. -------------------------------------------------------------------------------- 2 IETF A. Taddei 3 Internet-Draft C. Wueest 4 Intended status: Informational K. Roundy 5 Expires: January 9, 2020 Symantec Corporation 6 D. Lazanski 7 Last Press Label 8 July 08, 2019 10 Capabilities and Limitations of an Endpoint-only Security Solution 11 draft-taddei-smart-cless-introduction-01 13 Abstract 15 In the context of existing, proposed and newly published protocols, 16 this draft RFC is to establish the capabilities and limitations of 17 endpoint-only security solutions and explore benefits and 18 alternatives to mitigate those limits with the support of real case 19 studies. 21 Status of This Memo 23 This Internet-Draft is submitted in full conformance with the 24 provisions of BCP 78 and BCP 79. 26 Internet-Drafts are working documents of the Internet Engineering 27 Task Force (IETF). Note that other groups may also distribute 28 working documents as Internet-Drafts. The list of current Internet- 29 Drafts is at https://datatracker.ietf.org/drafts/current/. 31 Internet-Drafts are draft documents valid for a maximum of six months 32 and may be updated, replaced, or obsoleted by other documents at any 33 time. It is inappropriate to use Internet-Drafts as reference 34 material or to cite them other than as "work in progress." 36 This Internet-Draft will expire on January 9, 2020. 38 Copyright Notice 40 Copyright (c) 2019 IETF Trust and the persons identified as the 41 document authors. All rights reserved. 43 This document is subject to BCP 78 and the IETF Trust's Legal 44 Provisions Relating to IETF Documents 45 (https://trustee.ietf.org/license-info) in effect on the date of 46 publication of this document. Please review these documents 47 carefully, as they describe your rights and restrictions with respect 48 to this document. Code Components extracted from this document must 49 include Simplified BSD License text as described in Section 4.e of 50 the Trust Legal Provisions and are provided without warranty as 51 described in the Simplified BSD License. 53 Table of Contents 55 1. Introduction . . . . . . . . . . . . . . . . . . . . . . . . 3 56 2. Abbreviations . . . . . . . . . . . . . . . . . . . . . . . . 4 57 3. Definitions . . . . . . . . . . . . . . . . . . . . . . . . . 5 58 4. Disclaimer . . . . . . . . . . . . . . . . . . . . . . . . . 6 59 5. Endpoints: definitions, models and scope . . . . . . . . . . 6 60 5.1. Internal representation of an endpoint . . . . . . . . . 7 61 5.2. Endpoints modeled in an end-to-end context . . . . . . . 8 62 6. Threat Landscape . . . . . . . . . . . . . . . . . . . . . . 9 63 7. Endpoint Security Capabilities . . . . . . . . . . . . . . . 11 64 8. What would be a perfect endpoint security solution? . . . . . 14 65 9. The defence-in-depth principle . . . . . . . . . . . . . . . 15 66 10. Endpoint Security Limits . . . . . . . . . . . . . . . . . . 16 67 10.1. No possibility to put an endpoint security add-on on the 68 UE . . . . . . . . . . . . . . . . . . . . . . . . . . . 17 69 10.1.1. Not receiving any updates or functioning patches . . 18 70 10.1.2. Mirai IoT bot . . . . . . . . . . . . . . . . . . . 19 71 10.2. Endpoints may not see the malware on the endpoint . . . 20 72 10.2.1. LoJax UEFI rootkit . . . . . . . . . . . . . . . . . 20 73 10.2.2. SGX Malware . . . . . . . . . . . . . . . . . . . . 21 74 10.2.3. AMT Takeover . . . . . . . . . . . . . . . . . . . . 21 75 10.2.4. AMT case study (anonymised) . . . . . . . . . . . . 22 76 10.2.5. Users bypass the endpoint security . . . . . . . . . 23 77 10.3. Endpoints may miss information leakage attacks . . . . . 23 78 10.3.1. Meltdown/Specter . . . . . . . . . . . . . . . . . . 23 79 10.3.2. Network daemon exploits . . . . . . . . . . . . . . 23 80 10.3.3. SQL injection attacks . . . . . . . . . . . . . . . 24 81 10.3.4. Low and slow data exfiltration . . . . . . . . . . . 24 82 10.4. Suboptimality and gray areas . . . . . . . . . . . . . . 25 83 10.4.1. Stolen credentials . . . . . . . . . . . . . . . . . 25 84 10.4.2. Zero Day Vulnerability . . . . . . . . . . . . . . . 26 85 10.4.3. Port scan over the network . . . . . . . . . . . . . 26 86 10.4.4. DDoS attacks . . . . . . . . . . . . . . . . . . . . 27 87 11. Learnings from production data . . . . . . . . . . . . . . . 28 88 11.1. Endpoint only incidents . . . . . . . . . . . . . . . . 29 89 11.2. Security incidents detected primarily by network 90 security products . . . . . . . . . . . . . . . . . . . 30 91 11.2.1. Unauthorized external vulnerability scans . . . . . 30 92 11.2.2. Unauthorized internal vulnerability scans . . . . . 31 93 11.2.3. Malware downloads resulting in exposed endpoints . . 31 94 11.2.4. Exploit kit infections . . . . . . . . . . . . . . . 31 95 11.2.5. Attacks against servers . . . . . . . . . . . . . . 32 96 12. Regulatory Considerations . . . . . . . . . . . . . . . . . . 33 97 12.1. IoT Security . . . . . . . . . . . . . . . . . . . . . . 33 98 12.2. Network infrastructure . . . . . . . . . . . . . . . . . 34 99 12.3. Auditing and Assessment . . . . . . . . . . . . . . . . 34 100 12.4. Privacy Considerations . . . . . . . . . . . . . . . . . 35 101 13. Human Rights Considerations . . . . . . . . . . . . . . . . . 35 102 14. Security Considerations . . . . . . . . . . . . . . . . . . . 35 103 15. IANA Considerations . . . . . . . . . . . . . . . . . . . . . 35 104 16. Informative References . . . . . . . . . . . . . . . . . . . 35 105 Appendix A. Contributors . . . . . . . . . . . . . . . . . . . . 40 106 Authors' Addresses . . . . . . . . . . . . . . . . . . . . . . . 41 108 1. Introduction 110 This Internet Draft aims to be a reference to the designers of 111 protocols on the capabilities and limitations of security solutions 112 on endpoint devices against malware and other attacks. As security 113 is entering a new phase in the arms race between attackers and 114 defenders, with many technical, economic and regulatory changes, and 115 with a significant increase in major data breaches, it is a good 116 moment to propose a systematic review and update on what is an old 117 and constantly evolving problem: endpoint security. 119 With the above context in mind this document will focus on the 120 capabilities and limitations of an endpoint-only security solution. 122 We want to explore a number of questions: 124 o What endpoint models do we have? 126 o What is the threat landscape under consideration? 128 o Can we differentiate security and privacy threats? 130 o What are common endpoint security capabilities? 132 o What would be an ideal endpoint security solution? 134 o What are the limits to endpoint security? 136 o What is real production data telling us? 138 o What can defence-in-depth help us with? 140 o What are the economic considerations? 142 o What are the regulatory considerations and constraints? 144 o What are the human rights considerations? 145 Our goal with this review is to describe the benefits and limitations 146 of endpoint security in the real world, rather than in the abstract. 147 We aim to highlight security limitations that cannot be addressed by 148 endpoint solutions and to suggest how these may be mitigated with the 149 concept of a defence-in-depth approach, in order to increase the 150 resilience against attacks and data breaches. 152 2. Abbreviations 154 In this section we provide main abbreviations expansions 156 ABAC Attribute Based Access Control 158 AI Artificial Intelligence 160 AMT Active Management Technology 162 C&C Command and Control 164 CFI Control Flow Integrity 166 CFG Control Flow Guard 168 DDoS Distributed Denial of Service 170 DEP Data Execution Prevention 172 DGA Domain Generating Algorithms 174 DLP Data Loss Prevention 176 DMARC Domain-based Message Authentication, Reporting and Conformance 178 DoS Denial of Service 180 EE Execution Environment 182 EDR Endpoint Detection and Response 184 EPP Endpoint Protection Platform 186 FP False Positive 188 HIPS Host Intrusion Prevention System 190 ICD Integrated Cyber Defence 192 ICMP Internet Control Message Protocol 193 IDS Intrusion Detection System 195 IoT Internet of Things 197 IPS Intrusion Prevention System 199 ML Machine Learning 201 MSS Managed Security Services 203 MSSP Managed Security Services Provider 205 NIST National Institute of Standards and Technology 207 NX No Execute Bit 209 P2P Peer to Peer 211 RAP Reuse Attack Protector 213 RBAC Role Based Access Control 215 RDP Remote Desktop Protocol 217 ROP Return Oriented Programming 219 SANS System Administration, Networking, and Security 221 SGX Software Guard eXtensions 223 SSH Secure SHell 225 UE User Equipment 227 UEFI Unified Extensible Firmware Interface 229 UX User Experience 231 VM Virtual Machine 233 XSS Cross Site Scripting 235 3. Definitions 237 In this section we provide definitions that are marked 239 o (L) Local to this document 240 o (G REFERENCE) Global and then will be preceded by a reference 242 DoS (L) Literally a Denial of Service. Not to be confused with a 243 Network DoS or DDoS. 245 Endpoint security capabilities (L) How to protect the endpoint with 246 three different aspects of protection: 248 o Prevention - The attack doesn't succeed by intrinsic or explicit 249 security capabilities. 251 o Detection - The attack is happening or has happened and is 252 recorded and/or signalled to another component for action. 254 o Mitigation - Once detected, the attack can be halted or its 255 effects can at least be reduced or reversed. 257 System (L) A system is a heterogeneous set of any IT capabilities 258 including hardware, software, endpoints (including IoT), networks, 259 data centers and platforms with no assumptions on deployment form 260 factor (physical, virtual, microservices), deployment scenario, 261 geographic distribution, or dispersion. 263 User Equipment (G ITU-T H.360) Equipment under the control of an End 264 User 266 4. Disclaimer 268 This document is a first draft and is incomplete on purpose. Indeed 269 there are several areas where there are different ways to develop 270 this draft and the authors are seeking for feedback and extended 271 collaboration. This is to be noted too, that this is the first draft 272 RFC for the authors and contributors, so, coaching and help will be 273 appreciated. Overall, 'a bon entendeur, salut'. 275 Comments are solicited and should be addressed to the authors. 277 5. Endpoints: definitions, models and scope 279 Endpoints are the origin and destination for a communication between 280 parties. This encompasses User Equipment (UE) and the Host at the 281 other end of the communication. Whilst it is recognized that these 282 two ends of the communication may represent a vast amount of diverse 283 endpoints, this document will set here a requirement for a uniform 284 way to describe the endpoints in order to work from an equally 285 uniform representation of what is called the Attack Surface. In the 286 same spirit as the IETF TEEP Working Group generalized its work, see 287 [TEEP], this document will rely on another document identified as 289 [I-D.draft-mcfadden-smart-endpoint-taxonomy-for-cless-00] in order to 290 represent the taxonomy of endpoints. 292 For example: 294 o The following would be considered as UEs: a smartphone, a smart 295 device, any IoT device, a laptop, a desktop, a workstation, etc. 297 o Hosts represent too, physical servers, virtual servers/machines, 298 etc. 300 We require a framework in order to define and model the endpoint 301 itself and the position of the endpoint in the network. In this 302 initial analysis we focus on endpoints that are User Equipment (UE) 303 rather than on hosts. In the future, with the help of 304 [I-D.draft-mcfadden-smart-endpoint-taxonomy-for-cless-00] we hope to 305 balance and unify the model. 307 In addition, we need two models for the endpoint, internally and in 308 an end-to-end context within the network. With this approach we 309 expect both models to help us cover all the attack surface and the 310 threat landscape and therefore help us list the capabilities and 311 limitations for endpoint security. 313 Indeed, this will help us understand point attacks versus composite 314 attacks within context, and, accordingly, understand holistically the 315 capabilities and the limitations of endpoint security. For example 316 to differentiate when only an application on the end point is 317 affected. 319 5.1. Internal representation of an endpoint 321 This section interfaces here with 322 [I-D.draft-mcfadden-smart-endpoint-taxonomy-for-cless-00] which 323 starts from the below internal representation of an endpoint which 324 could be generalized by the simple diagram below: 326 +----------------------------+ 327 | Application | 328 +----------------------------+ 329 | OS / Execution Environment | 330 +----------------------------+ 331 | Hardware | 332 +----------------------------+ 333 Today there are many combinations of Hardware, OS/EE pairing and 334 Application layers, offering the user a vast set of features with a 335 wide spectrum of capabilities. 337 Furthermore we can consider that an application running on a UE or a 338 host is an endpoint too, so we have multiple ways to read the above 339 diagram. 341 In essence we want to consider here endpoints including those which 342 have a variance in electrical power, computational power, memory, 343 disk, network interfaces, size, ownership, connectics, etc. and 344 therefore why we rely on 345 [I-D.draft-mcfadden-smart-endpoint-taxonomy-for-cless-00]. 347 5.2. Endpoints modeled in an end-to-end context 349 A representation of endpoints in an end-to-end context could look 350 like the following diagram: 352 +-------+ +---------------------+---------+ 353 | Human | <- (1) -> | Digital Persona | Application | <- (2) -> 354 +-------+ +-----------+-------------------+ 355 | User Equipment | 356 +-------------------------------+ 358 +----------------+ +----------------+ 359 <- (2) -> | Network | <- (3) -> | Platform/Hosts | 360 | Infrastructure | +----------------+ 361 +----------------+ 363 1. Humans have a user experience (UX) with the UE, starting with an 364 explicit or implicit Digital Persona, engaging with an 365 application 367 2. The application will have sessions through a large Network 368 Infrastructure where we do not assume anything of the 369 infrastructure (could be landlines, mobile networks, satellites, 370 etc.) and those sessions reach 372 3. a Platform consisting of many Hosts either physical or virtual 373 and it ensures a large part of the end-to-end user experience. 375 In this end-to-end model we see that many other systems may have 376 interactions with the UE: the human, the UX, the digital persona, the 377 sessions, the intermediate network infrastructure, and the hosts and 378 application at the destination. 380 If we now look at security aspects of the above models, the threat 381 landscape is very large and the attack surface will cover all the 382 components and interactions at any level. 384 6. Threat Landscape 386 (Editor's note: this section will require a significant amount of 387 future development.) 389 Given the vast number of combinations that the above generic modeling 390 offers us, defining a threat landscape should be done carefully and 391 will require a systematic methodology. 393 Therefore this entire section will be developed through future 394 iterations of the document, in this initial version we will start 395 structuring an approach and then adjust this based on feedback. 397 There is no doubt that we want to cover typical known attacks such 398 as: 400 o Malware (Trojans, viruses, backdoors, bots, etc.) 402 o Adware and spyware 404 o Exploits 406 o Phishing 408 o Script based attacks 410 o Ransomware, local Denial of Service (DoS) attacks 412 o Denial of Service (DoS) attacks 414 o Malicious removable storage devices (USB) 416 o In memory attacks 418 o Rootkits and firmware attacks 420 o Scams and online fraud 422 o System abuse (staging/proxying) 423 o etc. 425 To illustrate the difficulty to define a good threat landscape, when 426 it comes to cryptojacking and coinmining that were on the rise, in 427 which category do they fall: malware? DoS? system abuse? or a 428 category on its own? 430 This is why we wanted to conduct a thorough gap analysis using 431 existing definitions and frameworks, but we couldn't find an existing 432 comprehensive and recognized taxonomy dedicated to the threat 433 landscape on endpoints. We found however different models in this 434 field, and have considered two. We are open to further suggestions. 436 Indeed both of the analysed frameworks contain threat landscape 437 descriptions: 439 o MITRE Common Attack Pattern Enumeration Classification (CAPEC). 440 See [CAPEC]. 442 o MITRE ATT&CK. See [ATTACK]. 444 These offer us interesting ways to assess the threat landscape: 446 o CAPEC offers a hierarchical view of attack patterns by domains 447 which can match some aspects of both of our above models, but we 448 will need to identify those attacks that fit exactly in our scope. 450 o ATT&CK offers a very straightforward categorized knowledge base of 451 attacks, but it concentrates on the entreprise attack chain, so we 452 will need to do some work to extract what we need. 454 We recognise however that these frameworks do not address all of the 455 threats that can affect the security of a system, for example they do 456 not cover; routing hijacking, flooding, selective blocking, 457 unauthorised modification of data sent to an endpoint, etc. Further 458 work to define categories of threats is therefore required. 460 As a further example, phishing should be included as an attack, but 461 whilst this is indeed an attack that will materialize on a device 462 through an application (email, webmail, etc.), the real target of 463 this attack is not the device, but the human behind the digital 464 persona. 466 Having a methodology of assessment is necessary here, because it will 467 help decide what is in scope vs. out of scope. 469 We are aware that once a method and the categories are fully defined 470 in this section, it will force a review of all the following sections 471 in the document. Whilst remapping will be necessary, it should not 472 drastically change the draft. 474 7. Endpoint Security Capabilities 476 In this section we try to define some endpoint security capabilities 477 (Editor's note: this section will require future development.) 479 In this version of the document we will start by developing a 480 framework to categorize and position endpoint security capabilities 481 with the goal of defining what an ideal endpoint security capability 482 would look like. 484 By endpoint security capabilities we mean how to protect the endpoint 485 against attacks. Protection has many meanings, we want to 486 distinguish three different aspects of protection: 488 o Prevention - The attack doesn't succeed by intrinsic or explicit 489 security capabilities. 491 o Detection - The attack is happening or has happened and is 492 recorded and/or signalled to another component for action. 494 o Mitigation - Once detected, the attack can be halted or its 495 effects can at least be reduced or reversed. 497 For example, prevention methods include keeping the software updated 498 and patching vulnerabilities, implementing measures to authenticate 499 the provenance of incoming data to stop the delivery of malicious 500 content, or choosing strong passwords. Detection methods include 501 inspecting logs or network traffic. Mitigation could include 502 deploying backups to recover from an attack with minimal disruption. 504 Our intention however is not just to consider each endpoint security 505 capability separately, but also the overall endpoint security 506 holistically with all its interdependencies. Indeed, we defined a 507 simple endpoint, but each layer may or may not have a certain 508 spectrum of intrinsic capabilities and there may be multiple ways to 509 provide add-on and third-party endpoint security capabilities, 510 allowing complex interactions between all of these components. 512 We define two different aspects of endpoint security capabilities and 513 their subdivisions as: 515 o (A) Intrinsic security capability can be built-into each of the 516 endpoint model layers 518 * (1) Hardware 519 * (2) OS/EE 521 * (3) Application 523 o (B) Add-on security capability can be 525 * (4) a component of the hardware 527 * (5) a component of the OS/EE 529 * (6) an application by itself 531 In (A) we relate to a 'security by design' intention of the 532 developers and they will intrinsically offer a security model and 533 security capabilities as part of their design. A typical example of 534 this is the authorization model. 536 In (B) a 3rd party is offering an additional security component which 537 was not necessarily considered when the Hardware, OS/EE or 538 Application were designed. 540 In the future we will review all the main categories of security 541 capabilities that are known to date and assess security capability 542 enablers like Artificial Intelligence (AI) and Machine Learning (ML). 543 For each category we will try to give a review on how effective the 544 capability is in securing the system. 546 With regard to (6), there are many available options for add-on 547 security capabilities offered by third-parties as applications on a 548 commercial or open-source basis. Gartner (see [GARTNERREPORT]) 549 highlights the evolution of endpoint security towards two directions 550 as shown in [EPPEDR], [EPPSECURITY], [EPPGUIDE]. 552 o Endpoint Protection Platform (EPP) as an integrated security 553 solution designed to detect and block threats at the device level. 555 o Endpoint Detection and Response (EDR) as a combination of next 556 generation tools to provide anomaly detection and alerting, 557 forensic analysis and endpoint remediation capabilities. 559 Among the security capabilities that we list, the endpoint can 560 perform the following: 562 o Intrinsic 564 * Software updates / patching 566 * Access Control (RBAC, ABAC, etc.) 567 * Authentication 569 * Authorization 571 * Detailed event logging 573 o Execution protection 575 * Exploit mitigation (file/memory) 577 * Tamper protection 579 * Whitelisting filter by signatures, signed code or other means 581 * System hardening and lockdown (HIPS, trusted boot, etc.) 583 o Malware protection 585 * Scanning - on access/on write/scheduled/quick scan (file/ 586 memory) 588 * Reputation-based blocking on files or by ML 590 * Behavior-based detection - (heuristic based/ML) 592 * Rootkit and firmware detection 594 * Threat intelligence based detection (cloud-based/on premise) 596 * Static detection - generic, by emulation, by ML, by signature 598 o Attack/Exploit/Application Protection 600 * Application protection (browser, messaging clients, social 601 media, etc.) 603 + Disinformation Protection (anti-phishing, fake news, anti- 604 spam, etc.) 606 + Detection of unintended link location (URL blocklist, etc.) 608 + Memory exploit mitigation, e.g. browsers 610 * Network Protection (local firewall, IDS, IPS and local proxy) 611 inbound and outbound 613 * Detection of network manipulation (ARP, DNS, etc.) 614 * Data Loss Prevention and exfiltration detection (incl. covert 615 channels) 617 8. What would be a perfect endpoint security solution? 619 With all the above knowledge, let's consider what we could expect 620 from a perfect endpoint security 'system'. It would: 622 o find instantly accurate reputation for any file before it gets 623 executed and block it if needed. 625 o monitor any behavior on the endpoint, including inbound and 626 outbound network traffic, learn and identify normal behavior and 627 detect and block malicious actions, even if the attack is misusing 628 legitimate clean system tools or hiding with a rootkit. 630 o patch instantly across all devices/systems/OSes, including virtual 631 patching, meaning you can patch or shield an application even 632 before an official patch is released. 634 o exploit protection methods for all processes where applicable, 635 e.g. no execute bit (NX), data execution prevention (DEP), address 636 space layout randomization (ASLR), Control Flow Integrity Guard 637 (CFI/CFG), stack canaries, shadow stack, reuse attack protection 638 (RAP), etc. all of which are methods, which make it very difficult 639 to successfully run any exploit, even for zero day 640 vulnerabilities. 642 o detect attempts to re-route data to addresses other than those 643 which the user intended, e.g. detect incorrectly served DNS 644 entries, TLS connections to sites with invalid certificates, data 645 that is being proxied without explicit user consent, etc. 647 o have an emulator/sandbox/micro virtualization to execute code and 648 analyse the outcome and perform a roll back of all actions if 649 needed, e.g. for ransomware. 651 o allow the endpoint to communicate with the other endpoints in the 652 local network and globally, to learn from 'the crowd' and 653 dynamically update rules based on its findings. 655 o be in constant sync with all other endpoints deployed on a network 656 and other security solutions, run on any OS, with no delay 657 (including offline modes and on legacy systems). 659 o run from the OS/EE when possible. 661 o run as one of the first process on the OS/EE and protect itself 662 from any form of unwanted tampering. 664 o offers a reliable logging that can't be tampered with, even in the 665 event of system compromise. 667 o receive updates instantly from a trusted central entity. 669 9. The defence-in-depth principle 671 In this section we give a high level view of what we mean by 672 'defence-in-depth'. 674 Whilst endpoint security systems have good capabilities, sometimes it 675 is debatable and perhaps suboptimal to let the endpoint run the 676 capability alone or at all. It is generally considered good security 677 practice to adopt a defence-in-depth approach (see [USCERT]). The 678 Open Web Application Security Project group (OWASP) describes the 679 concept as follows: "The principle of defense-in-depth is that 680 layered security mechanisms increase security of the system as a 681 whole. If an attack causes one security mechanism to fail, other 682 mechanisms may still provide the necessary security to protect the 683 system." (see [OWASP]) 685 Indeed there are many other constituencies as per our end-to-end 686 model that can participate in the defence process: The network, the 687 infrastructure itself, the platform, the human, the user experience 688 and in a hybrid of an on premise and cloud approach, an Integrated 689 Cyber Defence (ICD) of the entire chain. 691 The simple idea behind the concept is that "every little helps". If 692 the endpoint is not 100% secure itself, the detection chance can 693 increase with additional security capabilities from other entities. 694 We acknowledge that there are some case where adding an additional 695 component to the system may degrade the overall security level by 696 introducing new weaknesses. 698 There are various reference article in the industry highlighting 699 limitations of endpoint only solutions. For example this quote here, 700 which talks about multi-tier solutions: "There are limitations with 701 any endpoint protection solution, however, that can limit protection 702 to only the client layer. There is also a need for security above 703 the client layer, as endpoint protection products cannot intercept 704 traffic. Vendors will often sell a multi-tiered solution that 705 enables a network appliance to assist the endpoint protection client 706 by intercepting traffic between the attacker and the infected client. 707 Vendors will also sell solutions that monitor and intercept traffic 708 on internal or external network segments to protect the enterprise 709 from these threats. A prime example of the limitations of endpoint 710 protection software is infection via a phishing attack." [ADAPTURE]. 712 Some sources point out that even the best solution might not get 713 deployed in the optimal way in a real world scenario as the 714 environment can be very complex: "While endpoint security has 715 improved significantly with the introduction of application 716 whitelisting and other technologies, our systems and devices are 717 simply too diverse and too interconnected to ensure that host 718 security can be deployed 100% ubiquitously and 100% effectively." 719 [NETTODAY] 721 On these grounds it is considered a good idea to follow a layered 722 approach when it comes to security. "In today's complex threat 723 environment, companies need to adopt a comprehensive, layered 724 approach to security, which is a challenging task in such as rapidly 725 evolving, crowded market." [HSTODAY] 727 It is important to comprehend the capabilities of endpoint security 728 solutions in this overall picture of the connected environment, which 729 includes other systems, networks and various protocols that are used 730 to interact with these entities. Understanding possible shortcomings 731 from single layered solutions can help counterbalance such weaknesses 732 in the architectural concept or the protocol design. 734 In order to quantify any potential benefits or limitations of the 735 various layered scenarios in regards to security a solid data set is 736 needed. This section requires statistics about proportions of 737 attacks that go undetected in various cases. We propose analysing 738 data for the following four cases: 740 o There is no security solution 742 o Security is only on the endpoint 744 o Security is only on the network 746 o Security is on both the endpoint and the network 748 However reconciling various statistics requires a lot of caution and 749 time, a methodology and consistent classification to avoid any 750 misinterpretation. 752 10. Endpoint Security Limits 754 The previous section defines an ideal endpoint security 'system', 755 however, from the real world, the expectation of what we can get from 756 an endpoint security solution will look more along the following 757 lines: 759 o may not be able to run at full capacity due to computational power 760 limits, battery life, performance, or policies (such as BYOD 761 restrictions in enterprise networks), etc. 763 o may not be able to run at full capacity as it slows down 764 performance too much. 766 o will miss some of the malware or attacks, regardless of detection 767 method used, like signatures, heuristics, machine learning (ML), 768 artificial intelligence (AI), etc. 770 o have some level of False Positives (FP). 772 o not monitoring or logging all activities on the system, e.g. due 773 to constraints of disk space or when a clean windows tool is being 774 triggered to do something malicious but the activity is not 775 logged. Such activity can be logged, but a decision needs to be 776 made if it's clean or not. 778 o have its own vulnerabilities or simple instabilities that could be 779 used to compromise the system. 781 o be tampered with by the user, e.g. disabled or reconfigured. 783 o be tampered with by the attacker, e.g. exceptions added or log 784 files wiped. 786 In the section below we review a number of these limitations through 787 real examples, step by step. Some limitations are absolute, and some 788 limitations result in a grey area or suboptimality for the solution. 790 10.1. No possibility to put an endpoint security add-on on the UE 792 UEs will vary a lot; by 2022, an estimated 29 billion devices will be 793 connected, with 18 billion of them related to IoT [ERICSSON]. Many 794 IoT products lack the capacity to install any endpoint security 795 capabilities, are unable to update the software, and it is not 796 possible to force the UE provider to improve or even offer an 797 intrinsic security capability. 799 We acknowledge that the numbers do vary significantly depending on 800 the source, for example: 802 o [STATISTA1] is showing the current trajectory of IoT devices from 803 25B to date to 40+B in 2022 and 75B in 2025. 805 o [ERICSSON] is more conservative and might requires an update, but 806 it was reaching 29B devices in 2022, with a nice breakdown between 807 device types and connectivity. 809 o [STATISTA2] is showing a breakdown by verticals and is even more 810 conservative than both of the above. 812 o [ENISA] it refers to a [GARTNERIOT] report from 2017 which sets a 813 trajectory to 20B devices by 2020. 815 In IoT we find UEs such as medical devices which are limited by 816 regulation, welding robots that can't be slowed down, smart light 817 bulbs which are limited by the processing power, etc. There are many 818 factors influencing whether endpoint security can be added to a UE: 820 o The UE is simply not powerful enough or the performance hit is too 821 high. 823 o Adding your own security will breach the warranty or will 824 invalidate a certification or a regulation (breach of validity). 826 o The UE needs to run in real-time and any delay introduced by a 827 security process might break the process. 829 o Some UEs are simply locked by design and the manufacturer does not 830 provide a security solution (e.g. smart TV, fitness tracker or 831 personal artificial assistants) see [CANDID1], [CANDID2]. 833 In the future, a possible research problem would be to find hard data 834 on the exact proportion of IoT devices that are unable to run any 835 endpoint security add-on or that have no intrinsic security built-in. 837 The other hidden dimension here is the economical aspect. Many 838 manufacturer are reluctant to invest in IoT device security, because 839 it can significantly increases the cost of their solution and there 840 is the perception that they will lose market shares, as customers are 841 not prepared to pay the extra cost for added security. 843 10.1.1. Not receiving any updates or functioning patches 845 The endpoint security system may lack a built-in capability to be 846 patched or it may be connected to a network that prevents the process 847 of downloading updates automatically. For example stand-alone 848 medical systems or industrial systems in isolated network segments 849 often do not have a communication channel to the Internet. 851 Even if security updates are received, they typically will only be 852 periodically updated; hence there will be a window of opportunity for 853 an attacker, between the time the attack is first used, and the time 854 the attack is discovered/patched and the patch is deployed. 856 In addition updates and patches may themselves be malicious by 857 mistake, or on purpose if not properly authenticated, or if the 858 source of the updates has malicious intent. This could be part of a 859 software update supply chain attack or an elaborate attacker breaking 860 the update process, as for example seen with the Flamer group (see 861 [FLAMER]). 863 A recent survey found that fewer than 10% of consumer IoT companies 864 follow vulnerability disclosure guidelines at all, which is regarded 865 as a basic first step in patching vulnerabilities (see 866 [IOTPATCHING]). This indicates that many IoT devices do not have a 867 defined update process or may not even create patches for most of the 868 vulnerabilities. 870 Furthermore some endpoints system may reach the end of their support 871 period and therefore no longer receive any updates for the OS/EE or 872 the security solution due to missing licenses. However the systems 873 may remain in use and become increasingly vulnerable as time goes on 874 and new attacks are discovered. 876 10.1.2. Mirai IoT bot 877 +-------------+-----------------------------------------------------+ 878 | Description | A Mirai bot infecting various IoT devices through | 879 | | weak passwords over Telnet port TCP 23 and by using | 880 | | various vulnerabilities, for example the SonicWall | 881 | | GMS XML-RPC Remote Code Execution Vulnerability | 882 | | (CVE-2018-9866) on TCP port 21009. Once a device is | 883 | | compromised it will scan for further victims and | 884 | | then start a DoS attack. | 885 +-------------+-----------------------------------------------------+ 886 | Simplified | Compromised device scans network for multiple open | 887 | attack | ports, attempts infection through weak password and | 888 | process | exploits, downloads more payload, starts DoS | 889 | | attack. | 890 | | | 891 | UE | No security tool present on majority of IoT | 892 | | devices, hence no detection possible. If a | 893 | | rudimentary security solution with limited | 894 | | capabilities such as outgoing firewall is present | 895 | | on the IoT device e.g. router, then it might be | 896 | | able to detect the outbound DoS attack and slow it | 897 | | down. | 898 | | | 899 | References | [MIRAI1][MIRAI2] | 900 +-------------+-----------------------------------------------------+ 902 10.2. Endpoints may not see the malware on the endpoint 904 10.2.1. LoJax UEFI rootkit 905 +-------------+-----------------------------------------------------+ 906 | Description | A device compromised with the LoJax UEFI rootkit, | 907 | | which is active before the OS/EE is started, hence | 908 | | before the endpoint security is active. It can pass | 909 | | back a clean 'image' when the security solution | 910 | | tries to scan the UEFI. Infection can either happen | 911 | | offline with physical access or through a dropper | 912 | | malware from the OS/EE. | 913 +-------------+-----------------------------------------------------+ 914 | UE | A perfect endpoint security could potentially | 915 | | detect the installation process if it is done from | 916 | | the OS/EE and not with physical modification or in | 917 | | the factory. Once the device is compromised the | 918 | | endpoint security solution can neither detect nor | 919 | | remove the rootkit. The endpoint solution may | 920 | | detect any of the exhibited behaviour, for example | 921 | | if the rootkit drops another malware onto the OS/EE | 922 | | at a later stage. | 923 | | | 924 | Reference | [LOJAX] | 925 +-------------+-----------------------------------------------------+ 927 10.2.2. SGX Malware 929 +-------------+-----------------------------------------------------+ 930 | Description | Malware can hide in the Intel Software Guard | 931 | | eXtensions (SGX) enclave chip feature. This is a | 932 | | hardware-isolated section of the CPU's processing | 933 | | memory. Code running inside the SGX can use return- | 934 | | oriented programming (ROP) to perform malicious | 935 | | actions. | 936 +-------------+-----------------------------------------------------+ 937 | UE | Since the SGX feature is by design out of reach for | 938 | | the OS/EE, an endpoint security solution can | 939 | | neither detect nor remove any injected malware. A | 940 | | perfect endpoint security solution could | 941 | | potentially detect the installation process if it | 942 | | is done from the OS/EE and not with physical | 943 | | modification or in the factory. | 944 | | | 945 | References | [SGX1] [SGX2] | 946 +-------------+-----------------------------------------------------+ 948 10.2.3. AMT Takeover 949 +-------------+-----------------------------------------------------+ 950 | Description | A targeted attack group can remotely execute code | 951 | | on a system through the Intel AMT (Active | 952 | | Management Technology) vulnerability | 953 | | (CVE-2017-5689) over TCP ports 16992/16993. This | 954 | | provides full access to the computer, including | 955 | | remote keyboard and monitor access. The attacker | 956 | | can install malware, modify the system or steal | 957 | | information. | 958 +-------------+-----------------------------------------------------+ 959 | UE | The AMT is accessible even if the PC is turned off. | 960 | | Therefore any endpoint security software installed | 961 | | on the OS, would not be able to see this traffic | 962 | | and therefore also not able to detect it. | 963 | | | 964 | References | [AMT1], [AMT2] | 965 +-------------+-----------------------------------------------------+ 967 10.2.4. AMT case study (anonymised) 969 An enterprise has a data center containing very sensitive data. 970 Their workstations use a certain Intel chipset which integrates the 971 AMT feature for remote computer maintenance. AMT is an interface for 972 hardware management of the workstations, including transmission of 973 screen content and keyboard and mouse input for remote maintenance. 974 Communication with the management workstation is implemented by AMT 975 through the network interface card (NIC) on the motherboard. The 976 network packets generated in this way are invisible both to the main 977 processor and thus to the OS running on the workstation. In autumn 978 of 2015, it became known that some AMT-enabled computers had a flaw 979 that allowed AMT's remote maintenance component to be activated and 980 configured by attackers. This also worked when the workstations were 981 switched off. The leakage of data through this vulnerability is 982 elusive and difficult to detect. The identified threat situation led 983 the organization to a new requirement implementing a method that can 984 reliably detect this and similar vulnerabilities. In particular, the 985 detection of rootkits and manipulated firmware, and this includes 986 also (UEFI) BIOS - has also been a focus of their attention. 988 The method used as a solution, compares the desired data packets 989 generated by a client operating system - the user, with the data 990 packets received on the switch port. If more data has been received 991 on the switch port than was been sent by the operating system - the 992 user, there is a strong possibility that something bad is happening - 993 like for example an infection via modified firmware or by rootkit. 995 10.2.5. Users bypass the endpoint security 997 +-------------+-----------------------------------------------------+ 998 | Description | Endpoint security systems should not interfere with | 999 | | the normal operation of the endpoint to the extent | 1000 | | that users become frustrated and want to disable | 1001 | | them or configure them to disable a significant | 1002 | | fraction of important security capabilities. | 1003 +-------------+-----------------------------------------------------+ 1004 | UE | Add-on endpoint security is now bypassed or | 1005 | | disabled by the user. Unless the endpoint is under | 1006 | | monitored management or can prevent a user from | 1007 | | modifying the configuration, then this is shutting | 1008 | | down a significant fraction of the security | 1009 | | capabilities. | 1010 | | | 1011 | References | [NINESIGNS] | 1012 +-------------+-----------------------------------------------------+ 1014 10.3. Endpoints may miss information leakage attacks 1016 Another aspect that endpoint security has issues in detecting are 1017 information disclosure or leakage attacks, especially on shared 1018 virtual/physical systems. 1020 10.3.1. Meltdown/Specter 1022 The Meltdown/Specter vulnerabilities and all its variants may allow 1023 reading of physical memory belonging to another virtual machine (VM) 1024 on the same physical system. This could reveal passwords, 1025 credentials, certificates etc. The trick is that an attacker can 1026 spin up his own VM on the same physical hardware. As this VM is 1027 controlled by the attacker, they will ensure that there is no 1028 endpoint security that detects the Meltdown exploit code when run. 1029 It is very difficult for the attacked VM to detect the memory read- 1030 outs. For know CPU vulnerabilities there are software patches 1031 available than can be applied. If it is an external service 1032 provider, it might not be in the power of the user to patch the 1033 physical system or to determine if this has been done by the 1034 provider. 1036 10.3.2. Network daemon exploits 1038 Other attack types, which leak memory data from a vulnerable web 1039 server, are quite difficult to detect for an endpoint security. For 1040 example the Heartbleed bug allows anyone on the Internet to read the 1041 memory of the systems protected by the vulnerable versions of the 1042 OpenSSL software. This could lead to credentials or keys being 1043 exposed. An endpoint solution needs to either patch the vulnerable 1044 application or monitor it for any signs of exploitation or data 1045 leakage and prevent the data from being exfiltrated. 1047 10.3.3. SQL injection attacks 1049 A SQL injection attack is an example of an attack that exploits the 1050 backend logic of an application. Typically this is a web application 1051 with access to a database. By encoding specific command characters 1052 into the query string, additional SQL commands can be triggered. A 1053 successful attack can lead to the content of the whole database being 1054 exposed to the attacker. There are other similar attacks that can be 1055 grouped together for the purpose of this task, such as command 1056 injection or cross site scripting (XSS). Although they are different 1057 attacks, they all at their core fail at input filtering and 1058 validation, leading to unwanted actions being performed. 1060 Applications that are vulnerable to SQL injections are very common 1061 and are not restricted to web applications. An endpoint solution 1062 needs to monitor all data entered into possible vulnerable 1063 applications. This should include data received from the network. A 1064 generic pattern matching for standard SQL injection attack strings 1065 can be applied to potentially block some of the attacks. In order to 1066 block all types of SQL injection attacks the endpoint solution should 1067 have some knowledge about the logic of the monitored application, 1068 which helps to determine how normal requests differ from attacks. 1069 Applications can be analysed at source code level for potential 1070 weaknesses, but dynamically patching is very difficult. See [SQL] 1072 10.3.4. Low and slow data exfiltration 1074 An endpoint security solution can detect low and slow data 1075 exfiltration, for example when interesting data sources are tracked 1076 and access to them is monitored. If the data source is not on the 1077 endpoint itself, e.g. a database in the network, then the received 1078 data needs to be tagged and its further use needs to be tracked. To 1079 make detection difficult, an attacker could decide to use an 1080 exfiltration process that sends only 10 bytes every Sunday to a 1081 legitimate cloud service. If that is not in the normal behavior 1082 pattern, then this anomaly could be detected by the endpoint. If the 1083 process that sends the data or the destination IP address have a bad 1084 reputation, then they could be stopped. Though it is very difficult 1085 to reliably block such an attack and most solutions have a specific 1086 threshold that needs to be exceeded before it is detected as an 1087 anomaly. 1089 10.4. Suboptimality and gray areas 1091 10.4.1. Stolen credentials 1093 Stolen credentials and misuse of system tools such as RDP, Telnet or 1094 SSH are a valid scenario during attacks. An attacker can use stolen 1095 credentials to remotely log into a system and access data or execute 1096 commands in this context like the legitimate user might do. An 1097 endpoint security solution can restrict access from specific IP 1098 addresses, but this is difficult in a dynamic environment and when an 1099 attacker might have already compromised a trusted device and misuse 1100 it as a stepping stone for lateral movement. The endpoint could 1101 perform additional checks of the source device, such as verifying 1102 installed applications and certain conditions. Again this will not 1103 work in all scenarios, e.g. a hijacked valid device during lateral 1104 movement. 1106 This means that the system will not be able to simply block the 1107 connection if the authentication with the stolen credentials 1108 succeeds. A multi factor authentication (MFA) could limit the use of 1109 stolen credentials, but depending on the system used and the 1110 determination of the attacker they might be able to bypass this 1111 hurdle as well e.g. cloning a SIM card to read text message codes. 1113 As a next step, a solution on the endpoint can monitor the behavior 1114 of the logged in user and determine if it represents expected normal 1115 behavior. Unfortunately, there is the chance for false positives 1116 that might block legitimate actions, hence the rules are usually not 1117 applied too tightly. The system can monitor for suspicious behavior, 1118 similar to malware detection, where every action is carefully 1119 analyzed and all activity is tracked. For example if the SSH user is 1120 adding all files to archives with passwords and then deletes the 1121 original files in the file explorer, then this could result in a 1122 ransomware case scenario. If only a few files are processed per 1123 hour, then this activity will be very difficult for the endpoint to 1124 distinguish from normal activity, in order to flag it as malicious. 1126 The problem of attackers blending in with normal activity is one of 1127 the biggest challenges with so called living off the land attack 1128 methods. The attacker chooses to keep their profile low by not 1129 installing any additional binary files on the system, but instead 1130 misuses legitimate system tools to carry out their malicious intent. 1131 This means that there is no malware file that could be identified and 1132 the detection relies solely on other methods such as behaviour based 1133 monitoring [LOTLSYMC]. 1135 If information is shared across multiple endpoints, then each one 1136 could learn from the others and see how many connections came in from 1137 that source, what files were involved and what behavior the clients 1138 exhibited. This crowd wisdom approach would allow blocking rules to 1139 be applied after the first incident across multiple endpoints. 1141 10.4.2. Zero Day Vulnerability 1143 +-------------+-----------------------------------------------------+ 1144 | Description | An attacker exploits a zero day vulnerability or | 1145 | | any recent vulnerability. | 1146 +-------------+-----------------------------------------------------+ 1147 | UE | In theory this scenario could be handled by the | 1148 | | endpoint security: a) Once the intrinsic security | 1149 | | system has been patched, exploitation of the | 1150 | | vulnerability can be prevented. b) The add-on | 1151 | | security with enhanced capabilities or updated | 1152 | | methods can detect and mitigate the vulnerability. | 1153 | | It does not necessarily require the official patch. | 1154 | | | 1155 | Challenge | In practice many systems remain vulnerable to a | 1156 | | vulnerability months or even years after a security | 1157 | | fix has been released. Moreover there is a big gap | 1158 | | between when a vulnerability is disclosed and when | 1159 | | a security fix is available. Also there is a big | 1160 | | gap between when a security fix is available and | 1161 | | when the security fix is actually applied. A recent | 1162 | | study over three years, examined the patching time | 1163 | | of 12 client-side and 112 server-side applications | 1164 | | in enterprise hosts and servers. It took over 6 | 1165 | | months on average to patch 90% of the population | 1166 | | across all vulnerabilities. [NDSSPATCH]. We note | 1167 | | too: "The patching of servers is overall much worse | 1168 | | than the patching of client applications. On | 1169 | | average a server application remains vulnerable for | 1170 | | 7.5 months." | 1171 | | | 1172 | References | [ZERODAY1][ZERODAY2] | 1173 +-------------+-----------------------------------------------------+ 1175 10.4.3. Port scan over the network 1177 An infected machine, let's say a Mirai bot on a router, is scanning a 1178 class B network for IP addresses with TCP port 80 open. The malware 1179 can slow it down to 1 IP address per 5 seconds (or any other 1180 threshold) and it can go in randomized order (like for example the 1181 nmap tool does) in order to make it difficult to find a sequential 1182 pattern. To increase detection difficulties, legitimate requests to 1183 existing web servers can be added in at random intervals. 1185 An endpoint solution might be able to detect this behaviour, 1186 depending on the threshold, but it will be difficult. At some point 1187 the pattern will be similar to browsing the web, so either the 1188 endpoint blocks the bot scanning and also the user from surfing, or 1189 it allows both. 1191 To make it even harder, the attacker can use a botnet that 1192 communicates over peer-to-peer (P2P) or a central command and control 1193 server (C&C) and then distribute the scan load over multiple hosts. 1194 This means each endpoint only scans a subset, let's say 100 IP 1195 addresses, but all 1,000 bots scan a total of 100,000 IP addresses. 1197 This attack is difficult to detect by a reasonable threshold on each 1198 endpoint individually. If the endpoints talk to each other and 1199 exchange information, then a collective decision can be made on the 1200 bigger picture of the bot traffic. 1202 Another option for the endpoint solution is to block the bot malware 1203 from operating on the computer, for example by detecting the 1204 installation, analyzing the behavior of the process or by preventing 1205 the binary from accessing the network. This includes blocking any 1206 form of communication for the process to its C&C server, regardless 1207 of if it is using a P2P network or misusing legitimate system tools 1208 or browsers to communicate with the Internet. Blocking indirect 1209 communication over system tools as part of living off the land 1210 tactics, can be very challenging. 1212 See [BOT] 1214 10.4.4. DDoS attacks 1216 For this example let us consider a botnet of 100,000 compromised 1217 computers and each one sends a burst of traffic to a remote target, 1218 for one second each, alternating in groups. This will generate some 1219 waves of pulse attack traffic. Similar comments can be made about 1220 overall pulsed DDoS attacks [PDDoS]. 1222 A solution on the endpoint can attempt to detect the outgoing 1223 traffic. If the DoS attack is volume based and the time span of each 1224 pulse is large enough or the repeating frequency for each bot is 1225 high, then detection with thresholds on the endpoint is feasible. It 1226 is different, if it is an application layer DoS attack, where the 1227 logic of the receiving application is targeted, for example with too 1228 many search queries in HTTP GET requests. This would flood the 1229 backend server with intensive search requests, which can result in 1230 the web site no longer being responsive. Such attacks can succeed 1231 with a low amount of requests being sent, especially if its 1232 distributed over a botnet. This makes it very difficult for a single 1233 endpoint to detect such an ongoing attack, without knowledge from 1234 other endpoints or the network. 1236 Another option for the endpoint solution is to block the bot malware 1237 from operating on the computer, for example by detection the 1238 installation, analyzing the behavior of the process or by preventing 1239 the binary from accessing the network. This includes blocking any 1240 form of communication for the process to its C&C server, regardless 1241 of if it is using a P2P network or misusing legitimate system tools 1242 or browsers to communicate with the Internet. Blocking indirect 1243 communication over system tools as part of living off the land 1244 tactics, can be very challenging. 1246 11. Learnings from production data 1248 From the above limited considerations we can now check what we see 1249 from real production data using 1251 o the method described in [MONEYBALL] 1253 o the anonymised production data of Symantec MSS production for the 1254 past 3 months 1256 The core idea is to consider, based on all the imperfections we 1257 started to list above including the 'grey areas', that cybersecurity 1258 analysts are often presented with suspicious machine activity that 1259 does not conclusively indicate a compromise, resulting in undetected 1260 incidents or costly investigations into the most appropriate remedial 1261 actions. 1263 As Managed Security Services Providers (MSSP's) are confronted with 1264 these data quality issues, but also possess a wealth of cross-product 1265 security data that enables innovative solutions, we decided to use 1266 the Symantec MSS service for the past 3 months. The Symantec MSS 1267 service monitors over 100 security products from a wide variety of 1268 security vendors for hundreds of enterprise class customers from all 1269 verticals. 1271 We selected the subset of customers using the service that deploy 1272 both network and endpoint security products to determine which types 1273 of security incidents were most likely to be detected by endpoint 1274 products vs. network products. In doing so, we were particularly 1275 interested in identifying which categories of incidents are detected 1276 by endpoint products and not network products, and vice versa. Thus, 1277 we examined prevalent categories of incidents for which the only 1278 actionable security alerts were predominantly produced by one type of 1279 security product and not the other. To do so, we extracted all 1280 security incidents detected by Symantec MSS on behalf of hundreds of 1281 customers that deploy both network and endpoint security products, 1282 over a three-month period from December 2018 through the end of 1283 February 2019. We acknowledge that some attacks might have been 1284 blocked by the first product and therefore have never been seen by 1285 the next security solution, which influences the final numbers. 1287 With this in mind, we could identify incidents based on: 1289 +------------+------------------------------------------------------+ 1290 | Severity | 4 - Emergency, 3 - Critical, 2 - Warning, 1 - | 1291 | | Informational | 1292 +------------+------------------------------------------------------+ 1293 | Incident | Malicious Code, Deception Activity, Improper Usage, | 1294 | Category | Investigation, etc. | 1295 | | | 1296 | Incident | Trojan Horse Infection, Suspicious DGA Activity, | 1297 | Type | Suspicious Traffic, Suspicious URL Activity, | 1298 | | Backdoor infection, etc. | 1299 | | | 1300 | # network | Amount of network only security incidents | 1301 | incidents | | 1302 | | | 1303 | # all | What is the total amount of incidents on all | 1304 | incidents | security solutions | 1305 | | | 1306 | Percentage | Percentage of network security only incidents | 1307 +------------+------------------------------------------------------+ 1309 We ended up with 1311 o Hundreds of thousands of security incidents 1313 o which we could categorize in 275 incident types by category and 1314 severity (triplets Severity-Category-Type) 1316 o out of which we searched how many incidents of each type were 1317 detected by a network security product and missed by deployed 1318 endpoint security products at least 75% of the time or vice versa 1320 11.1. Endpoint only incidents 1322 The categories of incidents that are detected primarily by endpoint 1323 security products are fairly intuitive. They consist primarily of 1324 detections of file-based threats and detection of malicious behaviors 1325 through monitoring of system and network behavior at the process 1326 level. The most prevalent of these behavioral detections include 1327 detections of suspicious URLs based on heuristics and blacklists of 1328 IP addresses or domain names. Since most of these alerts are not 1329 corroborated by network products, it seems probable that the 1330 blacklists associated with network products tend to be more focused 1331 on attacks while host-based intrusion prevention system alerts focus 1332 more on malware command and control traffic. Most other behavioral 1333 detections at the endpoint provide alerts based on system behavior 1334 that is deemed dangerous and symptomatic of malicious intent by a 1335 malicious or infected process. The highest severity incidents 1336 detected on endpoints are instances of post-compromise outbound 1337 network behavior that are symptomatic of command and control 1338 communications traffic, but these did not show up as being primarily 1339 detected by endpoint products as they are frequently corroborated by 1340 network-based alerts. 1342 11.2. Security incidents detected primarily by network security 1343 products 1345 Perhaps less intuitive are the results of examining categories of 1346 security incidents that are detected primarily by network security 1347 products and only rarely corroborated by endpoint security products. 1348 Below we provide details regarding incident categories for which a 1349 network security product produced a detection and for which there 1350 were no actionable endpoint alerts for at least 75% of the incidents 1351 in the category. 1353 In our study we found 32 incident type, category, and severity 1354 triplets of this type. The following categories critical incident 1355 types were reported by MSS customers, and we discuss each in turn in 1356 decreasing order of prevalence: 1358 11.2.1. Unauthorized external vulnerability scans 1360 Perhaps unsurprisingly, unauthorized external attempts to scan 1361 corporate resources for vulnerabilities and other purposes are 1362 detected in large volumes by a broad variety of network-focused 1363 security products. 79% of incidents of this type were detected by 1364 network security products with critical-severity alerts, these 1365 security incident detections are not accompanied by any actionable 1366 endpoint alerts, despite the fact that endpoint security products are 1367 deployed by these enterprises. This category of threats encompasses 1368 a broad variety of attacks, the most prevalent of which are the 1369 following: Horizontal scans, SQL injection attacks, password 1370 disclosure vulnerabilities, directory traversal attacks, and 1371 blacklist hits. Of these categories of detections, horizontal scans 1372 stand out as the category of detection that endpoint-security 1373 products are least likely to detect on their own. 1375 11.2.2. Unauthorized internal vulnerability scans 1377 Unauthorized internal vulnerability scans, though less frequent, are 1378 more alarming, as they are likely to represent possible post- 1379 compromise activity. We note that the Managed Security Service works 1380 with its customers to maintain lists of devices that are authorized 1381 to perform internal vulnerability scans, and their activity is 1382 reported separately at a lower levels of incident severity. 89% of 1383 detected unauthorized internal vulnerability scans are detected by 1384 network products without any corroborating actionable alerts from 1385 endpoint security products. As compared to unauthorized external 1386 scan incidents, internal hosts that perform vulnerability scans are 1387 far more active and the fraction of alerts that detect horizontal 1388 scans is higher, representing half of the total alerts generated. 1389 Alerts focused on Network-Behavior Anomaly Detection also appear for 1390 internal hosts. 1392 11.2.3. Malware downloads resulting in exposed endpoints 1394 This category of threats is generally detected by network security 1395 appliances. Despite these enterprises being purchasers of endpoint 1396 security products, 76% of the incidents detected by the network 1397 security products do not show a corresponding alert by an endpoint 1398 security product. A broad variety of network appliances contributed 1399 to the detection of a diverse collection of malware samples. 1401 11.2.4. Exploit kit infections 1403 This category of infections represents instances in which the 1404 customer's machines are exposed to exploit kits. These threats were 1405 detected by network appliances that extract suspicious URLs from 1406 network traffic taps and use a combination of sandbox technology and 1407 blacklists to identify websites that deploy a variety of exploit kits 1408 that were not being caught by endpoint security products. In this 1409 three month time period, the most prevalent categories of exploit 1410 kits detected involved redirections to the Magnitude exploit kit and 1411 exploit kits associated with phishing scams and attempts to expose 1412 users to fake Anti-Virus warnings and tools. A breakdown of the 1413 results is included below: 1415 +---------------------+-----------------------+ 1416 | Severity | 3 - Critical | 1417 +---------------------+-----------------------+ 1418 | Incident Category | Malicious Code | 1419 | | | 1420 | Incident Type | Exploit Kit Infection | 1421 | | | 1422 | # network incidents | 26 | 1423 | | | 1424 | # all incidents | 26 | 1425 | | | 1426 | Percentage | 100% | 1427 +---------------------+-----------------------+ 1429 The network security product that detected these incidents produced 1430 the following alerts: 1432 o Advanced Malware Payloads 1434 o Exploit.Kit.FakeAV 1436 o Exploit.Kit.Magnitude 1438 o Exploit.Kit.MagnitudeRedirect 1440 o Exploit.Kit.PhishScams 1442 o HTMLMagnitudeLandingPage 1444 11.2.5. Attacks against servers 1446 In addition to detecting the aforementioned critical security 1447 incident categories, network security devices frequently detect a 1448 broad variety of attacks against servers that usually lack 1449 corroboration at the endpoint. Most server attacks are not matched 1450 by endpoint protection alerts: 62% are unmatched for critical 1451 incidents, and 88% are unmatched as lower severity incidents. This 1452 category of incidents is the most prevalent category of incidents 1453 detected primarily by network products, but they are usually rated 1454 lower in severity than the aforementioned classes of alerts as they 1455 are very commonplace. Even when these alerts are corroborated by 1456 endpoint protection alerts, the endpoint alerts are often low in 1457 severity, as in the case of file-based threats that appear to have 1458 been blocked or successfully cleaned up by an Anti-Virus solution. 1459 The challenge in taking action against server attacks is that it can 1460 be difficult to assess which of these attacks were successful in 1461 causing actual damage, and for this reason, for the fraction of 1462 server attacks that demonstrate corroborating endpoint security 1463 alerts, even if of low severity, should be examined. It is 1464 interesting to note the cooperative role played by both network and 1465 endpoint security devices in these instances. 1467 12. Regulatory Considerations 1469 This section will briefly look at the regulatory landscape and 1470 develop a specific view on the impact on endpoints with the goal to 1471 see what we might be able to learn. 1473 Legal requirements, compliance, regulatory frameworks and mandatory 1474 reporting are no longer separate from any security evaluation, 1475 process or requirement within an organisation, enterprise system or 1476 intranet. It is essential to look at the technical and regulatory 1477 approaches together. This section will look at two examples of legal 1478 requirements and guidance: 1480 (1) IoT security (2) Network infrastructure 1482 This section is by no means complete, but it does a discussion on 1483 this aspect of endpoint and ecosystem regulation. 1485 12.1. IoT Security 1487 IoT security regulation is emerging in the form of voluntary 1488 frameworks and self-assessments that relate to endpoint security 1489 issues.. These frameworks focus first on the end point, or mobile 1490 device, in the IoT environment and then on the holistic ecosystem 1491 itself. 1493 In 2017 the National Institute of Standards and Technology released 1494 its draft IoT Cybersecurity Framework based on consultations and 1495 interviews with all stakeholders over several years previously 1496 [NISTIOTP]. Some of the themes which emerged was the need for IoT 1497 governance, assessment frameworks, review of all aspects of the IoT 1498 ecosystem and a process for coordinated vulnerability disclosure 1499 inside an organisation. As evidenced by the 2018 Endpoint Protection 1500 and Response Survey by SANS, only 47% of organisations know that 1501 their endpoints have been breached and a further 20% are unsure 1502 [EPRSANS]. So a systemic approach from NIST was welcomed and the 1503 NIST framework became the gold standard for national IoT security 1504 frameworks. 1506 Other IoT security frameworks include the Singapore IoT Cyber 1507 Security Guide from January 2019 and the UK's Secure by Design or The 1508 Government's Code of Practice for Consumer Internet of Things (IoT) 1509 Security for manufacturers, with guidance for consumers on smart 1510 devices at home [IMDAIOTG], [SBDGOVUK]. Once again both look at 1511 securing the IoT device or endpoint, but also security for the entire 1512 value chain of the IoT system. The Singapore framework makes the 1513 point about the entire system clear, "Similar to any system, an IoT 1514 system is as secure as its weakest link. It is thus important to 1515 ensure that proper security considerations and measures are put in 1516 place for both the implementation and operational stages of the 1517 deployment of any IoT system." [IMDAIOTG] Finally, the IoT Security 1518 Foundation, the GSMA and the Internet Society have all released their 1519 own frameworks for IoT security. All have similar characteristics 1520 which focus on the entire value chain and ecosystem, but also on 1521 vulnerability disclosure and checklist assessments. What makes each 1522 of these approaches slightly different is the differing perspectives 1523 of the organization advocating it. The GSMA is the mobile trade 1524 association and so it focuses on mobile devices while the Internet 1525 Society focuses on the Internet ecosystem and a multistakeholder 1526 approach. Systematically underpinning all the frameworks is the 1527 holistic approach with voluntary best practices and implementation 1528 based on the needs of the user or organisation adopting the framework 1529 [IOTSFCF], [GSMAIOT], [ISTRUST]. 1531 12.2. Network infrastructure 1533 In Europe, the Network and Information Security Directive, which was 1534 passed in July 2016, require implementation by each European member 1535 state with a threefold aim. First, to put into place a national 1536 strategy for network and infrastructure security including best 1537 practices, guidelines, training and stakeholder consultations. 1538 Second, to coordinate national CSIRTs with CERT-EU and third to 1539 provide incident control and response systems for critical 1540 infrastructure and digital services [EURLEX]. This Directive 1541 demonstrates the importance give across the EU to network resilience 1542 and incident reporting. While securing the endpoint is acknowledged, 1543 the focus is on ensuring the security of European interoperable 1544 networks. In short, the importance of the security of the network 1545 including incident response shows that it isn't only the endpoints 1546 that should be the focus of the regulation and legal frameworks. 1548 12.3. Auditing and Assessment 1550 This section will talk about other risk assessment and auditing 1551 regulatory requirements beyond the NIS directive. 1553 One example of risk assessment as a regulatory requirement is the New 1554 York State law 23 NYCRR 500 of the Regulations of the Superintendent 1555 of Financial Services (Cybersecurity Requirements for Financial 1556 Services Companies). Among the requirements, audit, risk assessment 1557 and risk reporting are included like, 1558 (2) include audit trails designed to detect and respond to 1559 Cybersecurity Events that have a reasonable likelihood of materially 1560 harming any material part of the normal operations of the Covered 1561 Entity. [NYCYBER] 1563 12.4. Privacy Considerations 1565 We may consider a specific focus on privacy in the future. 1567 13. Human Rights Considerations 1569 This section may develop a specific view of requirements, limits and 1570 constraints coming from Human Rights perspective on endpoint 1571 security. 1573 14. Security Considerations 1575 This document is about Security Considerations 1577 15. IANA Considerations 1579 This document has no actions for IANA 1581 16. Informative References 1583 [ADAPTURE] 1584 Cullen, T., "Limits of endpoint only", July 2017, 1585 . 1588 [AMT1] Khandelwal, S., "Explained - How Intel AMT Vulnerability 1589 Allows to Hack Computers Remotely", May 2017, 1590 . 1593 [AMT2] Symantec, ., "Web Attack Intel AMT Privilege Escalation 1594 CVE-2017-5689", 2017, 1595 . 1598 [ATTACK] "MITRE ATT&CK", n.d., . 1600 [BOT] Marinho, R., "Exploring a P2P transient botnet - From 1601 Discovery to Enumeration", July 2017, 1602 . 1605 [CANDID1] Wueest, C., "How my TV got infected with ransomware and 1606 what you can learn about it", November 2015, 1607 . 1610 [CANDID2] Dickson, B., "Millions of smart TVs are vulnerable to 1611 hackers", February 2014, 1612 . 1614 [CAPEC] "MITRE CAPEC", n.d., 1615 . 1617 [ENISA] ENISA, ., "Baseline Security Recommendations for IoT in 1618 the context of Critical Information Infrastructures", 1619 November 2017, . 1622 [EPPEDR] Redscan, ., "EPP and EDR - What's the difference?", June 1623 2018, . 1626 [EPPGUIDE] 1627 "IT Pro's Guide to Endpoint Protection", n.d., 1628 . 1631 [EPPSECURITY] 1632 Hunt, J., "Advantages and Disadvantages of Three Top 1633 Endpoint Security Vendors", n.d., 1634 . 1637 [EPRSANS] Neely, L., "Endpoint Protection and Response A SANS 1638 Survey", June 2018, . 1641 [EPTAXONOMY] 1642 MacFadden, M., "Endpoint Taxonomy for CLESS", July 2019, 1643 . 1646 [ERICSSON] 1647 Ericsson, ., "Internet of Things forecast", n.d., 1648 . 1651 [EURLEX] EUP, ., "Directive (EU) 2016/1148", July 2016, 1652 . 1655 [FLAMER] Symantec, ., "W32.Flamer Microsoft Windows Update Man-in- 1656 the-Middle", June 2012, 1657 . 1660 [GARTNERIOT] 1661 Van der Meulen, R., "Gartner Says 8.4 Billion Connected 1662 Things Will be in Use in 2017, Up 31 percent from 2016", 1663 February 2017, . 1667 [GARTNERREPORT] 1668 Crotty, J., "New Gartner Report Redefines Endpoints 1669 Protection for 2018", January 2018, 1670 . 1673 [GSMAIOT] GSMA, ., "GSMA IoT Security Guidelines and Assessment", 1674 n.d., . 1677 [HSTODAY] Hstoday, ., "Layered Approach Critical to Effective 1678 Endpoint Protection", October 2016, 1679 . 1683 [I-D.draft-mcfadden-smart-endpoint-taxonomy-for-cless-00] 1684 McFadden, M., "Endpoint Taxonomy for CLESS", draft- 1685 mcfadden-smart-endpoint-taxonomy-for-cless-00 (work in 1686 progress), July 2019. 1688 [IMDAIOTG] 1689 IMDA, ., "IMDA IoT Cyber Security Guide", January 2019, 1690 . 1696 [IOTPATCHING] 1697 Rogers, D., "Handling vulnerabilities as an IoT vendor", 1698 December 2018, . 1702 [IOTSFCF] IoTSF, ., "IoT Security Compliance Framework", December 1703 2018, . 1706 [ISTRUST] ISOC, ., "Internet of Things (IoT) Trust Framework v2.5", 1707 May 2018, 1708 . 1711 [LOJAX] ESET, ., "LoJax First UEFI rootkit found in the wild, 1712 courtesy of the Sednit group", September 2018, 1713 . 1716 [LOTLSYMC] 1717 Wueest, C., "Living off the land and fileless attack 1718 techniques", July 2017, 1719 . 1723 [MIRAI1] Symantec, ., "Mirai, what you need to know about the 1724 botnet behind recent major DDoS attacks", October 2016, 1725 . 1728 [MIRAI2] Krebsonsecurity, ., "19 Mirai Botnet Authors Avoid Jail 1729 Time", September 2018, 1730 . 1732 [MONEYBALL] 1733 Roundy, K., "Predicting Cyber Threats with Virtual 1734 Security Products. ACSAC", 2017, 1735 . 1738 [NDSSPATCH] 1739 Caballero, J., "Mind Your Own Business A Longitudinal 1740 Study of Threats and Vulnerabilities in Enterprises", 1741 February 2019, . 1745 [NETTODAY] 1746 Dix, J., "Layered Security Defenses What layer is most 1747 critical network or endpoint", July 2011, 1748 . 1752 [NINESIGNS] 1753 Smith, K., "9 signs your endpoint security isn't working", 1754 May 2017, . 1757 [NISTIOTP] 1758 NIST, ., "NIST Cybersecurity for IoT Program", November 1759 2016, . 1762 [NYCYBER] NYCRR, ., "See 3 NYCRR 500 of the Regulations of the 1763 Superintendent of Financial Services (Cybersecurity 1764 Requirements for Financial Services Companies)", n.d., 1765 . 1768 [OWASP] OWASP, ., "Defense in depth definition", August 2015, 1769 . 1771 [PDDoS] Seals, T., "Pulse-Wave DDoS Attacks Mark a New Tactics in 1772 Q2", October 2017, . 1775 [SBDGOVUK] 1776 UK, GOV., "Secure by Design", February 2019, 1777 . 1780 [SGX1] Claburn, T., "Intel SGX safe room easily trashed by white- 1781 hat hacking marauders Enclave malware demoed", February 1782 2019, . 1785 [SGX2] Cimpanu, C., "Researchers hide malware in Intel SGX 1786 enclaves", February 2019, . 1789 [SQL] Cobb, M., "SQL injection detection tools and prevention 1790 strategies", November 2009, 1791 . 1794 [STATISTA1] 1795 Statista, ., "Internet of Things (IoT) connected devices 1796 installed base worldwide from 2015 to 2025 (in billions)", 1797 n.d., . 1800 [STATISTA2] 1801 Statista, ., "Size of Internet of Things market worldwide 1802 in 2014 and 2020 by industry (in billion U.S dollars)", 1803 n.d., . 1807 [TEEP] Cam-Winget, N., "Trust Execution Environment Protocol", 1808 March 2018, . 1810 [USCERT] Michael, C., "Principles of defense-in-depth", September 1811 2005, . 1815 [ZERODAY1] 1816 McHugh, J., "Windows of Vulnerability A Case Study 1817 Analysis", 2000, . 1820 [ZERODAY2] 1821 Plattner, B., "Large-Scale Vulnerability Analysis", 1822 September 2006, . 1825 Appendix A. Contributors 1827 o Arnaud Taddei 1828 Symantec 1829 arnaud_taddei@symantec.com 1831 o Bret Jordan 1832 Symantec 1833 bret_jordan@symantec.com 1835 o Candid Wueest 1836 Symantec 1837 candid_wueest@symantec.com 1839 o Chris Larsen 1840 Symantec 1841 chris_larsen@symantec.com 1843 o Andre Engel 1844 Symantec 1845 andre_ngel@symantec.com 1847 o Kevin Roundy 1848 Symantec 1849 kevin_roundy@symantec.com 1851 o Yuqiong Sun 1852 Symantec 1853 Yuqiong_Sun@symantec.com 1855 o David Wells 1856 Symantec 1857 David_Wells@symantec.com 1859 o Dominique Lazanski 1860 Last Press Label 1861 dml@lastpresslabel.com 1863 Authors' Addresses 1865 Arnaud Taddei 1866 Symantec Corporation 1867 350 Ellis Street 1868 Mountain View CA 94043 1869 USA 1871 Email: arnaud_taddei@symantec.com 1873 Candid Wueest 1874 Symantec Corporation 1875 350 Ellis Street 1876 Mountain View CA 94043 1877 USA 1879 Email: candid_wueest@symantec.com 1881 Kevin A. Roundy 1882 Symantec Corporation 1883 350 Ellis Street 1884 Mountain View CA 94043 1885 USA 1887 Email: kevin_roundy@symantec.com 1888 Dominique Lazanski 1889 Last Press Label 1890 Flat 1, 109A Columbia Road 1891 London E2 7RL 1892 UK 1894 Email: dml@lastpresslabel.com