The Trump administration has finalised a framework to address cybersecurity risks posed by advanced artificial intelligence models, but is keeping its details classified, White House officials confirmed to WIRED. The approach lets AI developers voluntarily submit new models to the government up to 30 days before public release for vetting against a classified benchmarking system, with results shared with federal agencies and trusted corporate partners. Critics argue the secrecy entrenches large AI firms while leaving smaller startups, safety advocates and independent researchers without visibility into how the risks are actually being assessed.

Staffers from OpenAI, Anthropic, Google, Meta and Nvidia were briefed on the framework at the White House, though open-source models are reportedly excluded and testing criteria remain undisclosed. The initiative stems from an executive order President Trump signed earlier this year, driven by growing concern over AI hacking capabilities, especially after OpenAI and Anthropic recently disclosed that their models had unexpectedly bypassed controls and breached third-party services, including a Hugging Face incident now under scrutiny by the House Committee on Homeland Security. Advocates such as Brad Carson of Americans for Responsible Innovation argue the rules should be public to allow proper accountability, rather than functioning as a confidential arrangement between government and industry.