CAISI ordered to stop public model evaluations amid new AI executive order

The Trump administration has classified AI safety evaluations after Anthropic's Mythos model exposed thousands of zero-day vulnerabilities in major operating systems.

Add us on Google by Editorial Team Jun. 9, 2026

The Center for AI Standards and Innovation, the US government body responsible for evaluating frontier AI models, has been told to stop publishing its findings. A new executive order signed on June 2, 2026, shifts control of AI model assessments from CAISI to a classified framework run by national security agencies.

The trigger: Anthropic’s unreleased Claude Mythos model, which autonomously identified thousands of previously unrecognized high-severity zero-day vulnerabilities in major operating systems.

What happened and why it matters

CAISI, formerly known as the US AI Safety Institute, had completed over 40 evaluations of AI models by early June 2026, building a public track record that gave researchers, policymakers, and the broader tech industry a shared understanding of what frontier models could actually do.

The new executive order mandates a 30-day government review of AI models before companies can release them publicly. The evaluation process is being handed to a trio of agencies: the Treasury Department, the NSA, and CISA (the Cybersecurity and Infrastructure Security Agency).

Anthropic announced the Mythos model in April 2026. Its autonomous cyber capabilities were apparently so advanced that the administration decided existing oversight mechanisms were insufficient.

CAISI had confirmed partnerships with Google DeepMind, Microsoft, and xAI as recently as May 5, 2026. Those relationships were built on the premise of open, collaborative evaluation. The new framework fundamentally changes that dynamic. Companies will still submit models for review, but the results will largely stay behind closed doors.

From open science to national security asset

The AI Safety Institute was originally established in 2022 to evaluate risk and safety in frontier models. Under the Trump administration, it was restructured and rebranded as CAISI, the Center for AI Standards and Innovation. Even under its new identity, CAISI continued publishing evaluations openly.

What this means for investors and the AI industry

The 30-day mandatory review period before any model release is a significant new friction point for every major AI company operating in the US. Anthropic, OpenAI, Google DeepMind, xAI, and every other frontier lab will need to build government compliance into their release schedules.

CAISI’s public evaluations served as a kind of neutral benchmarking service. With evaluations now classified, that independent verification disappears. Companies will tell you their models are great, but independent verification will no longer be publicly available.

US-based AI firms are now subject to a review process that Chinese, European, and other international competitors are not.

Researchers who joined CAISI to do open science now find themselves in a national security apparatus. Some will stay. Others will leave for academia or the private sector, where they can still publish.

Disclosure: This article was edited by Editorial Team. For more information on how we create and review content, see our Editorial Policy.