The White House has finalized a framework for reviewing cybersecurity risks from advanced AI models, but the public may not see the details. WIRED reports that the Trump administration briefed OpenAI, Anthropic, Google, Meta, Nvidia, and other AI companies on the plan this week.
Under the reported process, AI developers could voluntarily submit new models to the federal government up to 30 days before public release. The government would then evaluate cyber capabilities using a classified benchmarking system and share models with federal agencies and trusted corporate partners.
The secrecy is the central issue. WIRED reports that the administration is not disclosing the testing criteria or exactly which models will be covered, and Axios says open models are expected to be excluded. Critics argue that this could advantage large frontier labs while leaving smaller startups and outside researchers without clarity.
The policy reflects a real tension in AI security. Publishing cyber benchmarks can help accountability, but it can also reveal what attackers or model developers might optimize around. For now, companies and researchers outside the briefing room are being asked to trust a process they cannot fully inspect.