Observed Signal · Sep 15, 2026 · Policy Update · Source: Gary Marcus · Impact: 3/5 · Sentiment: Negative
Secret US AI evaluation framework partially revealed via FOIA
A US government framework for evaluating frontier AI models before release, previously secret, has been partially disclosed through a Freedom of Information Act (FOIA) request by Protect Democracy. The government released 132 pages of records, but most content is heavily redacted. The released pages confirm the involvement of top officials like Michael Kratsios (OSTP Director) and Ethan Klein (US CTO) but reveal few details about the evaluation criteria. Protect Democracy plans to continue litigation to seek full transparency. The framework is part of the administration's regulatory approach to AI, despite public statements against regulation.
The disclosure of a secret US AI evaluation framework is significant for AI companies and anyone deploying frontier models, as it affects regulatory clarity and compliance. The heavy redaction raises concerns about transparency, which could impact the AI industry's ability to plan and innovate.
Track Real-Time AI Regulation Signals & Market Shifts
Polaris7 autonomous intelligence agents track regulatory filings, primary sources, executive changes, and deal flow 24/7. Create your free Explorer workspace to monitor these entities.
Key Takeaways & Evidence Grounding
- Protect Democracy's FOIA request yielded 132 pages of records about the secret AI evaluation framework.
- The released documents are heavily redacted, revealing little about the framework's contents.
- Michael Kratsios and Ethan Klein are identified as participants in the framework's development.
- Protect Democracy plans to continue pushing for full disclosure through litigation.
- The framework is used to determine which 'frontier' AI models can be released.
Ontology Mapping & Concepts
Related Market Signals & Shifts
Recent verified developments and strategic activity across this market segment.
White House Hosts AI Firms to Review Model-Testing Framework
The White House will meet with leading artificial intelligence companies to review a newly completed voluntary framework for testing the cybersecurity capabilities of advanced AI models. Representatives from Anthropic are expected to participate, and OpenAI and Google are also reported to attend. The voluntary program would let participating developers give the government early access to certain frontier models for up to 30 days; the framework cannot be used to create mandatory federal licensing, permitting or preclearance requirements. The meeting follows a June 2 executive order directing federal agencies to create a process to identify “covered frontier models” and to develop classified benchmarks for assessing advanced cyber capabilities.
White House Dictates Access to Frontier AI Models
The Trump administration has started dictating which companies and entities can access the latest frontier AI models, shifting control previously held by developers like Anthropic and OpenAI. The White House launched a program called "Gold Eagle," a clearinghouse to collaborate with the private sector on cybersecurity vulnerabilities and to greenlight partner access to powerful models. The administration previously blocked access to Anthropic’s Mythos 5 and Fable 5 over national security concerns before reinstating access after negotiations; OpenAI limited new model access to "trusted partners" at the government's request. The moves follow a June executive order asking voluntary early access for government testing and come as international competitors such as China’s Moonshot AI narrow the performance gap.
Anthropic Influenced Trump AI Review Plan
A June 2026 analysis argues that Anthropic’s unpublished Mythos Preview model helped catalyze the Trump administration’s move toward a pre-release review process for powerful AI models. Initially reported as an executive order requiring formal government review, the White House revised the plan into a voluntary commitment that effectively pressures developers to submit models for classified review. The order directs the NSA to run a classified benchmarking process to identify “covered frontier models”; participating developers receive 30 days of government access and potential “trusted partner” status. The piece contends Anthropic’s public warnings about cybersecurity and recursive self-improvement align with its strategic interests—raising safety concerns while bolstering its market position ahead of an IPO—and warns the policy could create a two-tier system favoring large U.S. labs (Anthropic, Google, OpenAI) and reshape competitive dynamics.
Track Real-Time Market Signals & Shifts
Set up custom watchlists to receive automated, evidence-grounded executive digests whenever material signals or shifts occur across your tracked landscape.