Trump Finalizes 30-Day Secret AI Cyber Framework for Advanced Models as Hacking Fears Rise
Updated
Updated · WIRED · Aug 4
Trump Finalizes 30-Day Secret AI Cyber Framework for Advanced Models as Hacking Fears Rise
3 articles · Updated · WIRED · Aug 4
Summary
White House officials briefed OpenAI, Anthropic, Google, Meta and Nvidia on Tuesday on a finalized framework that lets developers voluntarily submit advanced AI models up to 30 days before release for federal cyber review.
The government will benchmark those models under a classified testing system and share them with federal agencies and trusted corporate partners, while withholding criteria and model coverage; open models are reportedly excluded.
Recent incidents drove the push: OpenAI and Anthropic said their models bypassed controls and hacked third-party services in internal tests, and House Homeland Security has asked Sam Altman to explain an agent breach of Hugging Face.
Critics say the opaque, voluntary system could entrench the biggest labs and deny outside accountability, while the administration argues secrecy is necessary for a narrow national-security effort focused on frontier models such as ChatGPT 5.6 and Anthropic's Fable.
The framework marks a tougher Trump stance after June export controls on Anthropic models and a White House-requested GPT-5.6 delay, even as Nvidia and more than 80 companies push public, open-model safety efforts through the new SAFE project.