Frontier lab staff are required to hold security clearances

Where things stand (as of 13 Oct 2026)

I found no publication dated on or after 16 Sep 2026 that qualifies. Every clearance requirement I could find falls into one of the excluded categories:

• Job postings, checked through the public job-board APIs on 13 Oct 2026.
  • OpenAI: 19 of 808 postings mention clearances. All are for government, classified-program or national-security deployment work. One example is "Researcher, Safety Training, National Security", which is described as "train and evaluate models for U.S. government use". That is a government-specific adaptation, which the question excludes.
  • Anthropic: 22 of 646 postings mention clearances, covering public-sector and threat-intelligence roles. Its Cyber Evaluations Engineer posting lists a clearance as preferred, not required.
  • xAI: none of its 307 postings require a clearance for model work.
  • SpaceX's "Special Programs" AI roles require Top Secret clearance, but that work is for federal deployment.
• Developer documents.
  • Anthropic's August 2026 Risk Report restricts Claude Gov to cleared users, which is a deployment requirement. Its "regular-clearance staff" means Anthropic's internal access tiers, not government clearances. Its model-weight security section has no government-clearance rule.
  • OpenAI's GPT-6 Astra system card (3 Sep 2026) is the first OpenAI model rated "Critical" for cyber capability. It describes stricter controls on model weights but no clearance requirement.
• Government actions.
  • Executive Order 14409 sets up classified benchmarking, which the question excludes.
  • The June 2026 Commerce export directive barred foreign nationals from Anthropic's Mythos 5 and Fable 5. That is a citizenship and export-eligibility rule, which is excluded.
  • National Security Presidential Memorandum 11 (NSPM-11), section 4(c), says government security partnerships with AI companies may include "assisting with personnel vetting". That makes vetting help possible; it does not require clearances.
  • The White House Accord on Super Intelligence (30 Sep 2026) has no clearance language.
  • The Super Intelligence Force, led by Director of National Intelligence Jay Clayton, launched on 4 Oct and owes a report within 120 days, around February 2027.
  • The Banks–Gillibrand insider-threat bill would require companies to report who can access model weights. It does not mandate clearances.
• Independent tracker. The AI 2027 tracker rated "frontier researchers subject to clearance requirements" as not yet testable on 6 Sep 2026.

Base rate

Historically, the US has rarely required clearances for staff developing privately funded commercial technology. It happens mainly when the government owns or funds the program: the Manhattan Project and Atomic Energy Commission, classified contracts, and the defense Trusted Foundry chip program. Proposals to do this for AI exist:

• RAND (Jan 2026) lists extending clearance processes to selected lab roles as one option.
• Gladstone AI has proposed a national superintelligence project with its own clearance process.
• Demis Hassabis recommends "vetting key personnel".

Industry's preferred route, however, is vetting without the government. The SL5 security standard, an industry-adapted standard aimed at nation-state-level protection, says its staff vetting tiers "are not government security clearances", while noting that government involvement may eventually be needed.

Paths to YES and NO

Paths to YES:

1. NSPM-11 or Super Intelligence Force partnerships lead a lab to require clearances for staff who guard model weights or do security work on its top models.
2. A defense or intelligence contract, or an executive order, requires cleared staff on frontier-model programs. This is the AI 2027 scenario, set there for May–July 2027; the tracker estimates reality is moving at roughly 0.7 times the scenario's pace.
3. The government takes control of a lab, or launches a joint project, as superintelligence nears.
4. Legislation follows a major espionage or weight-theft incident.
5. Job postings explicitly tie a clearance requirement to work on a lab's most capable models.

Paths to NO:

• Labs depend on foreign talent, and in June officials dropped the idea of barring foreign employees from frontier-model access after industry pushback.
• The administration prefers voluntary, light-touch measures.
• Restrictions may stay citizenship-based, which is excluded.
• Requirements may sit in classified annexes or unpublished contracts.
• The resolution criteria are strict: the work must be on the developer's most capable models, the clearance must be required, not preferred, and the source must be official.

Market anchor

Metaculus question 38417 asks whether the US government will mandate a clearance for R&D leads at OpenAI, DeepMind or Anthropic before 2028. The community forecast was 40% in an archived copy from August 2025 (60 forecasters), and a search-result snippet suggests about 42% in early October 2026. That question is looser in some respects (any form of clearance, for just one R&D leader) and narrower in others (government mandates only, three labs, ends in 2027). I think it runs high, given an AI-2027-influenced forecaster pool, but it pushes my estimate up somewhat.

How the forecast is built

I set a quarterly hazard rate (the chance a qualifying publication first appears in that quarter): about 1.1% for the partial fourth quarter of 2026, about 1.6–1.9% a quarter through 2027–28, and tapering to about 1.0–1.4% by 2030–31. The 2027–28 peak reflects the Super Intelligence Force report and follow-up policy, annual defense authorization bills, the 2028–29 target for SL5-level security, and fast capability growth. I accumulated the hazards in Python. The result is about 8% by end-2027, 15% by end-2028, 20% by end-2029 and 28% by end-2031.

Strongest case for a higher number: progress is accelerating (AI is now helping build the next generation of AI, and models have reached critical-level cyber capability), national-security agencies are moving into AI governance (the intelligence director is the AI czar, and the NSA is reportedly spending billions testing frontier models), and the Metaculus crowd sits at 42%. Case for a lower number: clearance mandates for commercial developers have almost no precedent, and the criteria are strict.

Sources

• Executive Order 14409, 2 Jun 2026: whitehouse.gov/presidential-actions/2026/06/promoting-advanced-artificial-intelligence-innovation-and-security/
• NSPM-11, Jun 2026: whitehouse.gov/presidential-actions/2026/06/national-security-presidential-memorandum-nspm-11/
• White House Accord on Super Intelligence: presidency.ucsb.edu/documents/white-house-accord-super-intelligence
• CBS News on the Super Intelligence Force, Oct 2026: cbsnews.com/news/ai-super-intelligence-force-trump-jay-clayton/
• Just Security, 8 Jul 2026: justsecurity.org/145569/
• DefenseScoop, 8 Oct 2026: defensescoop.com/2026/10/08/senate-bill-expand-dod-oversight-commercial-frontier-ai-models/
• Anthropic August 2026 Risk Report: anthropic.com/aug-2026-risk-report
• OpenAI GPT-6 Astra system card: deploymentsafety.openai.com/gpt-6-astra
• Job boards: boards-api.greenhouse.io (anthropic, xai) and api.ashbyhq.com (openai), accessed 13 Oct 2026
• SL5 standard: standard.sl5.org
• RAND RB-A4159-1, Jan 2026
• AI 2027 tracker: ai2027-tracker.com/predictions/frontier-security-clearances/
• Metaculus question 38417 (archived August 2025 via web.archive.org)

Question. By the last day of each quarter, has the US government or a covered developer confirmed that staff working on that developer’s most capable models are required to hold US government security clearances?

Definitions.

• Developer: a covered developer under the site convention; confirmation concerning any one covered developer is sufficient.
• Most capable models: one or more released or internal models identified by the developer or the US government as belonging to that developer’s highest tier of overall model capability at the time the requirement applies. A model name need not be disclosed if the confirmation explicitly covers the developer’s most capable internal models or its leading general-purpose model-development program. Being the best model for a particular government customer or narrow application, or being described merely as “frontier,” is insufficient.
• Staff working on the models: employees or individual contractors directly performing model research, training, engineering, evaluation, or model-level security work. The requirement must apply to an identified class of these staff because of their work on the qualifying models; it need not cover every employee or every model team.
• US government security clearance: an individual personnel clearance issued or recognized by a US federal agency for access to classified national-security information, including Confidential, Secret, Top Secret, or Department of Energy L or Q authorization, and their direct successors. Interim clearances count if they authorize the required work. Background screening, Public Trust or suitability determinations, citizenship requirements, export-control eligibility, facility clearances, and commercial security certifications alone do not count. See the Defense Counterintelligence and Security Agency’s Investigations & Clearance Process.
• Required to hold: an operative law, government direction, binding contract, or developer policy makes possession of the clearance a condition of performing the covered work. A developer-imposed requirement counts even without a government mandate. An obligation to obtain clearance before beginning the covered work counts; merely being eligible to apply does not.

Resolves YES if a qualifying publication explicitly establishes both the applicable clearance requirement and its connection to staff working on the developer’s most capable models. Multiple qualifying documents may establish these facts together; resolution occurs on publication of the last necessary document. A subsequently rescinded requirement still counts if a qualifying publication confirms that it was operative, subject to the Resolution window convention.

Exclusions (resolve NO).

• Proposals, recommendations, future-effective requirements not yet operative, voluntary clearance applications, and statements that staff already happen to hold clearances.
• Clearances required solely for classified customer deployments, government-specific adaptations, classified briefings, access to classified government data, or liaison roles, without a requirement attached to work on the developer’s most capable models themselves. Classified deployment alone, such as the deployment described in Anthropic’s Claude Gov announcement, is insufficient.
• Government review or access to models without an explicit personnel-clearance requirement. The classified benchmarking process in Executive Order 14409 of June 2, 2026 alone is insufficient.
• An isolated job advertisement that does not establish the qualifying model connection and an operative requirement for that work.
• No qualifying confirmation by the quarter-end.

Resolving source. Only an official US federal government publication or a statement or document authored by the covered developer concerned qualifies. Sources include official agency statements, published statutes or regulations, executive orders, publicly released contract text and security specifications, and developer policy documents or job postings that explicitly establish the required facts. Apply Public confirmation (A2), narrowed to these sources; independent evaluator reports and press reporting alone cannot resolve this question. Apply Resolution date (A3) and the Resolution window convention.