The Center for AI Standards and Innovation (CAISI) announced on 5 May that it had signed agreements with Google DeepMind, Microsoft and xAI to evaluate their frontier AI models before public release. The announcement extended existing arrangements with OpenAI and Anthropic and represented the broadest US government pre-deployment evaluation programme to date. Within days, the announcement page disappeared, with no public explanation offered for the removal.
Disclosure: RECATOOLS is written with the assistance of Claude, made by Anthropic, which is also named in this article. Readers should weigh that interest.
The episode adds to a run of stalled and restarted initiatives in US federal AI policy since the change in administration in early 2025. Pre-deployment evaluation, in which government tests frontier models in confidential environments before they reach customers, has bipartisan support, yet its implementation has lurched repeatedly from one approach to another.
What was announced
CAISI's 5 May announcement extended its existing pre-deployment evaluation framework to three additional labs. Under the agreements, Google DeepMind, Microsoft and xAI give the US government early access to their models to assess the systems' capabilities and help improve their security before public release. CAISI said it would "conduct pre-deployment evaluations and targeted research to better assess frontier AI capabilities and advance the state of AI security". It also said it had completed more than 40 evaluations of AI models since 2024, including state-of-the-art models that remain unreleased, and that OpenAI and Anthropic had renegotiated their existing partnerships with the centre.
The framework is voluntary. Labs are not legally required to participate, and nothing reported about the agreements gives CAISI's findings the power to block a release. The arrangement is closer to the UK's AI Safety Institute model — collaboration on capability discovery rather than approval-gate regulation — than to the EU AI Act's structured pre-deployment requirements.
Why this is unusual
The least expected name on the list is xAI. Elon Musk's lab had previously been outside the existing CAISI arrangements, and its decision to take part suggests its interests now line up with the federal evaluation regime in a way the company's earlier public stance did not predict.
The timing also stands out. The announcement landed in a month when reports said Anthropic's run-rate revenue had exceeded $44 billion and that it was raising at least $30 billion, and only weeks after the Mythos model controversy raised regulator attention to frontier offensive capability. A voluntary framework like CAISI's lets federal agencies stay current with the capability frontier without imposing a heavier regulatory hand.
Then the page came down
Within days of the announcement, CAISI's web page describing the new agreements was taken offline. The Commerce Department, under which CAISI sits, did not explain the removal, and whether the agreements themselves remain in effect has not been publicly confirmed.
One possibility is internal disagreement in the executive branch about how much pre-deployment evaluation should be made public, driven by concern that publicising the framework gives competitors and adversaries a roadmap of US capability assessment. Another is friction over which body owns the framework, between CAISI and other parts of the policy apparatus such as White House AI offices, the National Security Council and individual cabinet departments. The simplest explanation is that the page went live before it had been fully cleared, and the takedown is the routine consequence.
None of the three explanations has been confirmed, and we have found no public statement from CAISI or the labs on whether the arrangements continue unchanged.
The capability the framework is trying to evaluate
Pre-deployment evaluation matters most for capabilities where a model going public produces irreversible consequences. The policy literature concentrates on offensive cyber capability, biological-threat uplift and autonomous-agent behaviour.
Of these, offensive cyber capability is the most acute. The Mythos model is the public-facing example: Anthropic has described a frontier model capable of identifying chains of low-severity bugs that compose into high-severity exploits, and decided to gate access through Project Glasswing rather than release the model openly. The CAISI framework gives federal evaluators visibility into similar capabilities at Google DeepMind, Microsoft and xAI before the labs make release decisions.
On biological threats, research published over the past year has shown that frontier models can provide non-trivial uplift to actors attempting to synthesise dangerous biological agents, though the actual magnitude of uplift remains contested.
With autonomous agents the concern is different. As models gain longer-horizon planning capability and broader tool access, evaluators are looking for evidence of behaviour that diverges from operator intent — situational awareness, deception, self-preservation behaviours. These are softer evaluations with less mature methodology, but they are formally in scope.
How the labs are positioning
The labs probably take part because voluntary participation forestalls compulsory regulation. If the federal government can credibly say it has visibility into capability development through a framework that does not block releases, the political pressure for stricter, slower regimes is reduced. Because the federal government is a major customer for cloud AI services, the framework's findings can also shape future procurement decisions, which adds a commercial logic on top of the political one.
What to watch
Whether the framework survives its messy rollout will show first in whether the CAISI page returns to the public web, and in what its language says if it does. The other early sign is whether other major labs (Meta, the Chinese labs operating in US markets, mid-tier players like Cohere and Mistral via their US subsidiaries) sign on or stay out. The clearest test would be findings from any specific model evaluation becoming public, even at the level of capability category, since that would show the framework doing real work.
At the centre of the US policy debate is whether voluntary pre-deployment evaluation is the right durable answer for frontier AI. Supporters argue that the alternative (compulsory licensing, EU-style structured assessment, or moratoriums) imposes costs out of proportion to the marginal safety gain, while critics counter that voluntary frameworks lose force the moment a lab decides participation is no longer in its interest. How capability develops over the next 12 months is likely to settle which side has the stronger case.
Comparisons with the UK AISI and EU AI Office
| Body | Jurisdiction | Posture | Findings publication |
|---|---|---|---|
| CAISI | United States | Voluntary | Largely private |
| UK AI Safety Institute | United Kingdom | Voluntary | Selective public methodology |
| EU AI Office | European Union | Binding under AI Act | Mandated incident reports |
CAISI's framework sits in a small cluster of comparable bodies that have emerged in major jurisdictions over the past two years. The UK AI Safety Institute, established under the previous government and continued under the current one, has been operating a similar voluntary pre-deployment evaluation arrangement with OpenAI, Anthropic, Google DeepMind and a handful of other labs. The institute has published parts of its evaluation methodology and a small subset of its findings, building a public-facing reputation that lends weight to its private assessments.
The EU AI Office, the regulatory body created to enforce the AI Act, sits at the opposite end of the spectrum. Where CAISI and the UK AISI rely on voluntary participation and informal findings, the EU AI Office operates within a binding legal framework that requires general-purpose AI providers to meet specific transparency, documentation and risk-management obligations. The provisions for general-purpose AI models with systemic risk — a category broad enough to cover every frontier model in current deployment — include mandated incident reporting, evaluation cooperation and post-market monitoring.
The three regimes are not mutually exclusive. The same lab can participate in CAISI's evaluations, the UK AISI's evaluations and the EU AI Office's compliance regime simultaneously. In practice this is what is happening. The labs maintain three separate evaluation streams running in parallel, with the operational cost of doing so absorbed as a category of regulatory overhead.
Whether the three regimes converge or diverge over the next 18 months is one of the more consequential questions in international AI policy. If they settle on shared evaluation methodology, interoperable findings and mutual recognition of capability assessments, the labs' operational costs would fall and everyone would get more usable safety data. If three methodologies instead produce three different conclusions about the same models, all three regimes would lose credibility, and calls would grow for either international coordination or unilateral action by whichever regime moved most aggressively. The politics of that coordination remain fragile.
Correction, 18 October 2026: An earlier version said CAISI staff had been told to remove the announcement page, that the labs and CAISI had said the agreements remained in effect, and that the three labs had commented on the voluntary nature of the deal; we found no reporting to support those statements. We have also removed an unsupported description of Microsoft's coverage, replaced an unsourced account of what the evaluations cover with CAISI's own description, and corrected the reported Anthropic revenue figures.
Sources
View author profile → · Editorial policy
