Prolepsia forecastYes / no
Will NIST publish final security guidance for AI agents before 31 December 2027?
As of 18 November 2026, Prolepsia puts the chance of yes at 92%.
92%chance of yes
How Prolepsia reasons
The report Prolepsia wrote with this forecast.
How forecasts are made · Track record
TL;DR
Probability of Yes: 91.6%. Probability of No: 8.4%. The likeliest path is a final version of NIST’s adversarial machine-learning taxonomy that keeps its concrete agent safeguards; the question does not require a standalone agent security standard.
Key evidence
- The strongest candidate already contains actionable guidance. NIST published the initial draft of SP 2400-102e2026 on November 9, 2026, with comments due December 18. Section 4 describes checking proposed agent actions against policy before execution. That goes beyond naming a threat or listing a mitigation. The 2023 edition went from draft to final in 302 days; this edition has 416 days from draft release to the cutoff.
- NIST has more than one path to a qualifying publication. The preliminary Cyber AI Profile contains agent-specific access and authorization provisions. NIST is also developing agent control overlays and an agent identity implementation. The October 16 joint incident guide came out before the forecast window, but shows that NIST can co-issue practical guidance on a narrow agent-security topic.
- Two smaller paths add to the case. The TEVV-Athlon draft briefly covers testing agents for tool misuse and unauthorized actions. A November 12 Senate appropriations proposal calls for CISA agent identity guidance. That is not an enacted NIST deadline, and a CISA-only release would not count.
Uncertainties
The taxonomy remains a draft as of November 18, 2026, with no confirmed final date. It could take another review round or lose the implementation detail that makes it a strong candidate. The backup projects could face the same review or staffing delays. Only an official final or equivalent release with actionable agent-specific guidance, publicly available by December 30, 2027, at 11:59 p.m. U.S. Eastern Time, counts.
Read the full report
TL;DR
Probability of Yes: 91.6%. Probability of No: 8.4%. Watch NIST’s adversarial machine-learning taxonomy first: the decisive test is whether NIST publishes it by December 30, 2027, with its concrete agent safeguards intact.
Context
The forecast is dated November 18, 2026. A YES requires at least one official NIST or NIST-co-issued document to be made public in final or equivalently authoritative form by December 30, 2027, at 11:59 p.m. U.S. Eastern Time. It needs concrete cybersecurity guidance specific to AI agents. An actionable section in a broader document is enough. A draft, a passing mention, or a document first made public on December 31 does not count.
NIST moved from a formal request for information in January 2026 to an Agent Standards Initiative announced February 17. Its May 18 analysis of responses found broad agreement that existing cybersecurity practices need to adapt to agents. Those steps establish a work program, not a publication deadline. Some NIST materials use “SI” in place of “AI” under the September 29 executive order. What matters here is the kind of system the guidance covers, not the label.
Evidence
The broad publication record argues against assuming that a public draft will go final within a year. Among CSRC initial public drafts with a full 377 days of follow-up after comments closed, 20 of 32 from 2023, 19 of 37 from 2024, and 19 of 30 from 2025 reached final within that span. That is 58 of 99 overall, or 58.6%. In a separate CSRC catalogue comparison, 30 of 48 formal Special Publication drafts outside the SP 1800 practice-guide series reached final within 416 days, or 62.5%. The two comparisons use different clocks and publication mixes. They measure whether a document went final, not whether it contained guidance that meets this question’s criteria.
The closest precedent is more favorable. The 2023 edition of the adversarial machine-learning taxonomy reached final in 302 days from draft release and 96 days from comment close. The 2025 edition was published on March 24, 2025, but has no listed public draft from which to measure a lag. The 2026 draft appeared on November 9, with comments due December 18. That leaves 416 days from the draft and 377 from the scheduled comment close. This is an established publication series with enough time to match its closest observable precedent, not a guarantee that it will do so.
The strongest evidence is in the draft itself. Section 4 of SP 2400-102e2026 describes checking an agent’s proposed actions against policy before execution. It also discusses monitoring sequences of actions and limiting privileges in multi-agent systems. Those are concrete safeguards, not just labels for types of attacks. A final version with comparable detail would likely qualify. A final version that reduces the same material to a summary of the literature might not.
Two other NIST projects have a clear fit on substance. The preliminary Cyber AI Profile, IR 8596, was released on December 16, 2025, and includes provisions on agent authorization and permissions. It still has a formal draft stage ahead of final. The control overlays are a more direct fit because the planned agent volume covers both single- and multi-agent systems. It is last in the proposed sequence, and the planning materials do not establish a firm release date for a final volume.
The NCCoE agent identity project has moved beyond a concept paper: its September 29 update identified an existing DevSecOps project as the home for the first agent implementation. That makes a practical release more plausible, but completing an implementation is not the same thing as issuing final guidance. Incident response offers a different path. On October 16, 2026, NIST, NSA, and CISA jointly issued practical guidance for incidents during training and evaluation, including incidents involving agents. The guide predates the qualifying window. Its value is as a precedent for a narrow joint release. NIST’s November workshop report identifies follow-on work, but gives no dated release commitment.
What's non-obvious
The title of the leading document matters less than what it tells readers to do. The taxonomy draft does not need to become a practice guide to qualify; it needs to retain implementation-oriented recommendations. The same distinction opens a less obvious path through evaluation. Appendix B of the August 2026 TEVV-Athlon draft addresses how to test agents for tool misuse and unauthorized actions. The treatment is brief, so it is a modest backup. But concrete security assessment procedures can qualify even if the publication’s main subject is evaluation.
The policy signal is real but indirect. The November 12 Senate appropriations proposal calls for CISA to publish agent identity guidance within 180 days of enactment. It does not set an enacted NIST deadline. A CISA-only document would not count. Prior joint identity-security work makes NIST co-issuance plausible, but it is an additional route, not a commitment.
Uncertainties
The main uncertainties are timing and content. The taxonomy’s publication record shows an initial draft, not a confirmed final date. Another substantial round of review could push it past the cutoff. A timely final could still fall short if it drops the agent-specific safeguards or describes them without implementation guidance. The strongest update would be a revised Section 4 that retains the safeguards, paired with a dated final publication milestone.
The fallback projects do not eliminate the deadline risk. Several can be delayed by the same review queues, staffing constraints, or changes in priorities. A formal public draft of the agent overlay, a dated incident-response companion draft, or a publication schedule for the identity practice guide would clarify which backups can finish in time. A workshop, announcement, or live project page can show progress, but it cannot substitute for the publicly accessible final document that resolves the question.
Forecast history
| Date | Forecast |
|---|---|
| 92% chance of yes |
Sources Prolepsia read
- Federal Register :: Request for Information Regarding Security Considerations for Artificial Intelligence Agentswww.federalregister.gov
- Federal Register / Vol. 91, No. 190 / Friday, October 2, 2026 / Presidential Documents — 63129www.whitehouse.gov
- Search | CSRCcsrc.nist.gov
- NCCoE agent identity projectwww.nist.gov
- The TEVV-Athlon Framework for Evaluating AI Systems - ipdnvlpubs.nist.gov
- U:\2027REPT\15REPT\HSEXPLSTMT.TXTwww.appropriations.senate.gov
- New Concept Paper on Identity and Authority of Software Agents | NCCoEwww.nccoe.nist.gov
The question
This question asks whether the United States National Institute of Standards and Technology (NIST) will officially publish at least one final document providing substantive cybersecurity guidance specifically applicable to artificial intelligence (AI) agents before December 31, 2027. The qualifying publication must be publicly available no later than December 30, 2027, at 11:59 p.m. U.S. Eastern Time. AI agents are AI-based systems capable of independently planning, making decisions, using tools, interacting with external systems, or executing actions to accomplish specified objectives. Their ability to take autonomous actions creates cybersecurity risks involving unauthorized tool use, prompt injection, identity and access management, privilege escalation, data exposure, and interactions among multiple agents. As of November 18, 2026, NIST has several initiatives addressing these risks. On January 12, 2026, NIST's Center for AI Standards and Innovation (CAISI) issued a Request for Information concerning the security of AI agent systems. On February 17, 2026, NIST announced its AI Agent Standards Initiative, which includes research into agent security and identity and the development of guidelines supporting secure agent adoption. On May 18, 2026, NIST published an analysis of responses to its security-focused request for information, identifying widespread agreement that AI agents introduce security challenges requiring adaptations to existing cybersecurity practices. NIST is also developing Control Overlays for Securing AI Systems (COSAiS), including separate use cases covering single-agent and multi-agent systems. Most recently, on November 9, 2026, NIST released an initial public draft of Special Publication 2400-102e2026, Adversarial Machine Learning: A Taxonomy and Terminology of Attacks and Mitigations. This draft includes a taxonomy of attacks against agentic systems and corresponding mitigations, with public comments due December 18, 2026. Although this represents substantive progress, it has not yet been finalized. Furthermore, its primary purpose is to establish a taxonomy rather than provide comprehensive operational security guidance. Relevant official sources include NIST's AI Agent Standards Initiative (https://www.nist.gov/artificial-intelligence/ai-agent-standards-initiative), NIST's security and resilience research page (https://www.nist.gov/artificial-intelligence/ai-research-security-and-resilience), and the November 2026 draft publication (https://csrc.nist.gov/pubs/sp/2400-102/e2026/ipd). The forecast concerns whether NIST will progress from its ongoing research, consultations, and draft publications to at least one publicly released, finalized document containing actionable security guidance for AI agents within the specified deadline. It does not require NIST to produce a comprehensive security standard covering every aspect of agent security.
How it resolves
The question resolves YES if NIST officially publishes at least one qualifying final security guidance document between November 18, 2026, and December 30, 2027, inclusive. A qualifying document must satisfy all of the following conditions. First, it must be officially issued by NIST or one of its organizational units, including its Information Technology Laboratory, Computer Security Division, Center for AI Standards and Innovation, or National Cybersecurity Center of Excellence. Second, it must be publicly available through an official NIST website, the NIST Computer Security Resource Center, the NIST publications database, or another official NIST publication channel. Third, it must be identified as a final publication, final guideline, completed practice guide, or equivalent authoritative release rather than a preliminary draft, initial public draft, request for comments, working paper, or consultation document. Fourth, it must explicitly address the cybersecurity of AI agents or agentic AI systems and provide substantive, actionable recommendations, controls, safeguards, implementation practices, or security assessment procedures specifically applicable to such systems. Qualifying security guidance may address topics including agent identity and authentication, authorization and permission management, secure tool execution, prompt injection mitigation, protection of agent memory and contextual information, inter-agent communications, monitoring and auditing, or preventing unauthorized autonomous actions. The document need not cover every topic, but it must contain concrete security recommendations that go beyond merely describing threats or defining terminology. A finalized NIST Special Publication, NIST Interagency/Internal Report, Cybersecurity White Paper, National Cybersecurity Center of Excellence practice guide, or comparable official guidance document may qualify, regardless of its publication series. A final version of NIST's Control Overlays for Securing AI Systems addressing single-agent or multi-agent security would qualify if it contains substantive security recommendations for those systems. A final version of NIST SP 2400-102e2026 would qualify only if its finalized content includes actionable security mitigation guidance specifically addressing AI agents; a taxonomy that merely identifies and categorizes attacks and mitigations without implementation-oriented recommendations would not qualify. A document focused primarily on AI security more broadly may qualify if it contains a clearly identifiable section or set of controls specifically addressing security risks arising from AI agents. Merely mentioning AI agents as an example within general AI security guidance is insufficient. The question resolves NO if no qualifying final document has been officially published by December 30, 2027, at 11:59 p.m. U.S. Eastern Time. It also resolves NO if NIST has announced forthcoming guidance, released only drafts, conducted consultations, or published research identifying security threats without issuing qualifying final guidance by the deadline. Resolution will rely primarily on NIST's official publications database (https://www.nist.gov/publications), the Computer Security Resource Center publications directory (https://csrc.nist.gov/publications), and official NIST initiative and project pages. The publication itself and its official publication status will take precedence over secondary reporting. If a document's publication date is disputed, contemporaneous official NIST publication announcements, dated document records, and archived official publication pages will be used to establish whether it was publicly released before the cutoff. If no qualifying publication can be verified from official NIST sources after a reasonable verification period, the question resolves NO rather than being annulled.
Fine print
The deadline is interpreted strictly: 'before 31 December 2027' means publication must occur no later than December 30, 2027, at 11:59 p.m. U.S. Eastern Time. A document first published on December 31, 2027, does not qualify, even if published before the end of that day. Publication means that the final document is publicly accessible, not merely approved internally, submitted for publication, announced as forthcoming, or circulated among stakeholders. A document dated before the deadline but demonstrably made public afterward does not qualify. The document does not need to be legally binding, mandatory for federal agencies, or adopted by industry. Official voluntary NIST security guidance is sufficient. A jointly authored or jointly published document qualifies if NIST is explicitly identified as an issuing organization and the document meets the other criteria. Guidance published solely by another U.S. agency, an international standards organization, or an industry consortium does not qualify, even if NIST contributed to its development. Revisions or finalizations of documents originally drafted before November 18, 2026, qualify if the final version is first published during the forecast period. Existing final documents that are merely referenced, republished without substantive changes, or highlighted in new announcements do not qualify. A document that is officially designated final by NIST qualifies even if NIST subsequently publishes corrections, updates, or replacement editions. Conversely, documents explicitly identified as public drafts, preliminary guidance, or consultation materials do not qualify merely because they contain detailed security recommendations. If NIST undergoes an organizational restructuring or changes the names of its relevant divisions, publications issued by its direct organizational successor may qualify, provided the issuing body remains an official component of NIST or its legally designated successor. The determination will be based on official government documentation.
More public forecasts
Will the European Commission open a Digital Markets Act case about AI agent access to a gatekeeper's platform before 31 December 2027?
68%chance of yes
Prolepsia forecast
What will be the highest success rate on ExploitGym's full 869-task suite reported for any AI model or agent system by 30 June 2027?
70.3%success rate, any model, median
Prolepsia forecast
Will Los Angeles host the 2028 Summer Olympics with no event moved or cancelled for security reasons?
96%chance of yes
Prolepsia forecast
Will US securities class action filings exceed 230 in calendar year 2027?
49%chance of yes
Prolepsia forecast
How many CVE records will NIST's National Vulnerability Database list as published in calendar year 2027?
Prolepsia forecast
How many CVE records will NIST's National Vulnerability Database list as published from 1 January to 31 March 2027?
69,500CVE records, median
Prolepsia forecast