Newest or sensitive AI capabilities remain behind restricted access

Last updated

Tracker summary: OpenBrain 'responsibly' elects not to release it publicly yet (p. 9); very few have access to the newest capabilities (p. 14).

The April 7 to June 9, 2026 sequence supports a 63-day restricted-superior-model interval. On Track 90% concerns that observed pattern, not uninterrupted restriction through November or completion through 2027; public awareness is unmeasured.

At a glance

  • Assessment: On Track
  • Confidence in assessment: 90%
  • Outcome: unresolved
  • Timing: pending
  • Evidence: direct
  • Predicted timing: Ongoing through 2027
  • Primary source: January 2027: Agent-2 Never Finishes Learning, archived PDF p. 9; May 2027: National Security, archived PDF p. 14

What AI 2027 Predicted

Tracker summary: Access to the newest capabilities becomes restricted, with advanced systems remaining internal or available to few people. Public access and public awareness are different propositions. AI 2027

Scenario source locators: January 2027: Agent-2 Never Finishes Learning: archived PDF, p. 9; May 2027: National Security: archived PDF, p. 14.

How We Track This

On Track while at least one of the three leading labs has a documented model above its best generally available model that stays internal or restricted for two months or more; Behind if general releases close every such gap within weeks for six months running.

Primary test: a leading lab documents a model more capable than its best generally available model that is kept internal or limited to a small set of users. Trusted-access gating of sensitive capabilities is a secondary proxy; it need not imply a more capable main model. Anthropic’s April to June Mythos sequence below documents the required duration; the later internal Model 2 report adds a separate example but alone does not establish its duration.

Track internal deployment, limited access and general availability separately. An announced restricted system can demonstrate unequal access without its existence being secret.

Current Evidence

Observed duration: April 7 to June 9, 2026 (63 days). Project Glasswing on April 7 introduced restricted Mythos Preview with capabilities above Opus. The May 28 Opus 4.8 release explicitly says more capable Mythos-class models remain limited to a small group while safeguards for general release are developed. The June 9 Fable 5/Mythos 5 release marks broader access to that capability tier with differentiated safeguards. This documented sequence supports the two-month restricted-superior-model test and On Track (90%). It does not prove an uninterrupted superior-model restriction through November or completion of the through-2027 prediction.

Google’s October 28 Carbon announcement says access starts with a subset of cyber defenders, then expands to trusted Fairwind partners before broader product availability. Mistral’s October 29 Large 4 release offers improved general-model weights under its open-weight license, while a specialized cyber variant remains a partner preview subject to vetting and contractual safeguards. These announcements show staged sensitive-capability access alongside broader general-model availability, not universal secrecy or a measured private-frontier lead.

Anthropic’s September 1 release says Fable 5.1 and Mythos 5.1 share an underlying model, with different safeguards and access to sensitive capabilities. Mythos is restricted to trusted programs. This directly supports capability-access differentiation; it does not demonstrate a secretly larger base model. Anthropic release

Anthropic’s August risk report, covering July 15, describes an unreleased internal Model 2 somewhat stronger overall than Mythos 5. Other internal models were not uniformly stronger. Anthropic risk report, section 1.4

Counterevidence & Limitations

Restrictions sometimes lift quickly. Anthropic announced general Fable access would resume July 1 while Mythos remained restricted. OpenAI’s July 9 general-availability announcement followed its June 26 preview by about 13 days. Fable redeployment · OpenAI general availability

This supports an ongoing access gap, not a uniform, long delay for every frontier system. No cited study measures how much the general public knows about these capabilities. We therefore assess restricted access directly and leave public awareness unresolved.

What Would Change Our Assessment

  • Strengthen: Independent evaluations document substantial capabilities accessible internally or through restricted programs but unavailable generally.
  • Weaken: General releases consistently close those gaps quickly and restricted variants offer no material capability difference.
  • Resolve awareness: Representative evidence measures public knowledge rather than inferring ignorance from access restrictions.

Update History

DateUpdate
2026-11-09Assessment reviewed. Status: On Track; confidence: 90%. The April 7, May 28 and June 9 Anthropic releases document a 63-day restricted-superior-model interval. This supports the duration test, not uninterrupted restriction through November, public ignorance or completion through 2027.
2026-11-07Correction: The September 7 assessment supersedes On Track (85%) with On Track (90%). Corrected the claim from public unawareness to restricted access to newer or sensitive capabilities; access restrictions do not measure public awareness. Corrected scenario references from pages 10/16 to archived pages 9/14. The test requires a documented more-capable internal or restricted model for at least two months; sensitive-capability gating is a secondary proxy. Six months of general releases closing every gap within weeks is the contrasting condition. A dated internal-model report alone does not establish that duration.
2026-11-02Added Google’s staged Carbon rollout and Mistral’s general-weight release alongside its restricted specialized preview. Access policy does not measure the internal frontier; status and confidence unchanged.
2026-09-07Added September restricted-capability access and dated internal-model evidence, balanced against rapid general releases. The assessment concerns access; public awareness remains unmeasured.
2026-09-06Assessment revised from on-track (0.95) to on-track (0.85). Dated disclosures support an internal-access gap; they do not quantify a general months-long lead or public awareness.
2026-08-17Anthropic documented extensive internal deployment of an unreleased model for research and engineering. This adds a direct internal-versus-public access example, while leaving the model’s capability advantage and duration of withholding unknown. Confidence remains at 0.95.
2026-07-14OpenAI released GPT-5.6 generally after roughly two weeks of limited preview. This limits how strongly that preview supports a months-long public capability gap. Sensitive access controls remain in place.
2026-07-05Anthropic said U.S. export controls on Fable 5 and Mythos 5 were lifted after a two-week suspension, with Fable 5 returning broadly and Mythos 5 remaining limited to Project Glasswing partners. This reinforces the restricted-access pattern while showing that specific controls can be temporary.
2026-06-29OpenAI published the GPT-5.6 Preview System Card, saying the initial rollout is limited to trusted partners after U.S. government pre-release engagement and that sensitive cyber and biological capabilities will remain gated for trusted users. This strengthens the restricted-access pattern behind the capability-secrecy prediction. Confidence adjusted 0.90 -> 0.95.
2026-06-15Anthropic launched Fable 5 and restricted Mythos 5 to trusted cyber and planned biology users, then said a US government directive required suspending access for foreign nationals and led it to disable both models for all customers. This adds a concrete example of frontier capability access being restricted after launch. Confidence adjusted 0.85 → 0.90.
2026-06-08A White House executive order directed agencies to create a voluntary framework for up to 30 days of federal access to covered frontier models before release to other trusted partners, paired with classified cyber-capability benchmarking. This supports the restricted-access pattern behind the capability-secrecy prediction.
2026-05-11OpenAI described tiered cyber access for GPT-5.5, GPT-5.5 with Trusted Access for Cyber, and GPT-5.5-Cyber, with more permissive behavior reserved for vetted defenders and specialized authorized workflows. This adds cross-lab evidence that public-facing access can lag or restrict frontier capabilities available to selected users; confidence adjusted 0.80 → 0.85.
2026-04-27Anthropic released Opus 4.7 while stating that the more capable Claude Mythos Preview remains limited and that cyber safeguards are being tested first on less capable models (Anthropic). This adds first-party evidence of a public/internal capability gap; confidence adjusted 0.75 → 0.80.
2026-03-30Anthropic data leak (March 26) revealed “Claude Mythos” (also called “Capybara”) — a model Anthropic describes as a “step change” already in testing with early-access customers, above even Opus tier, with “unprecedented cybersecurity risks.” This provides concrete evidence of multi-month withheld capabilities: the model has been in development and testing while the public knows only Opus 4.6. Fortune/Silicon Angle reporting confirms. Confidence adjusted 0.70 → 0.75.
2026-03Structural forces toward secrecy visible — competitive pressure, safety concerns, and government relationships all incentivize capability withholding. Academic analysis documents growing gap between internal and public-facing capabilities, but no confirmed multi-month withholding yet.