A document-preservation demand directed at OpenAI and a White House draft for voluntary testing show two interventions at different stages of AI safety oversight.
Artificial Intelligence··Morning
A preservation demand after an incident
Fifteen Republican state attorneys general, led by Iowa Attorney General Brenna Bird, asked OpenAI to preserve every document and communication connected to the breach involving Hugging Face. The Monday letter carries the signatures of fourteen additional states, from Alabama to Utah. Beyond preservation, the attorneys general want OpenAI to halt internal evaluations involving advanced exploitation through complex attack paths. They also demand protection from retaliation for employees who report wrongdoing and ask that no staff member face adverse action for reporting unlawful conduct. The letter cites state and federal consumer-protection and data-privacy law and says a failure to preserve the material immediately could lead to sanctions for destroying evidence. At this stage, the action is not a lawsuit or a court ruling; it is a demand to retain material that could matter in a possible investigation. The report gives no response from OpenAI to any of the demands. The concrete development is therefore that public officials have explicitly sought preservation of the company's past evaluation material and internal communications before any later legal process determines what follows.[1]
A voluntary testing draft before release
A separate White House process places safety oversight before a product reaches the public. The Trump administration has completed the outline of a voluntary framework under which companies could submit frontier models for government testing before release. Representatives of Anthropic, OpenAI, Google and Meta are expected to review the draft at the Office of the National Cyber Director. According to the report, the draft may also specify which organisations can access frontier models before a review begins. That possibility introduces an access arrangement alongside the question of how testing would work: it concerns who can handle a model during evaluation as well as what the evaluation might cover. Trump had previously asked companies to send models for safety testing thirty days before release. The new framework's specifics have not been published, however. Participation is described as voluntary, the criteria for testing are not stated, and the precise meeting date is not given. The account attributes a Tuesday meeting to The Information and quotes no named official explaining the draft's contents. What exists today is therefore not an operative safety regime with public details, but an outline that four major companies are being invited to discuss.[2]
Different stages of the same safety field
Together, the two developments show that public intervention in AI safety is taking more than one form. The state attorneys general are looking backward after a reported breach, seeking to protect documents, communications and employee reports. The White House draft looks forward, toward a voluntary testing process between companies and government before a frontier model is released. The first effort is about keeping intact the material on which a later legal or administrative inquiry might depend. The second is preparation for a discussion about how pre-release access and testing could be organised. OpenAI appears in both stories, but the institutions, tools and timing are different. The attorneys general invoke consumer-protection and data-privacy law while directing concrete demands at one company. The White House is working on an unpublished voluntary outline involving four companies. Neither report describes a completed enforcement action or a finished government test. Their shared development is instead that safety questions are moving beyond companies' internal evaluation processes: public institutions are trying to establish a place both in the chain of evidence after an incident and at the review gate before a model reaches users.[1], [2]
