White House Hides AI Cyber Framework From Public
Get the Tech newsletter
Daily tech — startups, AI labs, chips, the launches that shape the next decade. Free.
- The White House finalized a classified AI cybersecurity framework after inviting OpenAI, Anthropic, Google, Meta, and Nvidia to a Tuesday briefing, allowing developers to voluntarily submit models up to 30 days before public release for vetting by a classified benchmarking system.
- The framework excludes open-weight models and hides its testing criteria, prompting critics like Americans for Responsible Innovation president Brad Carson to call it "a cloak of secrecy" that amounts to "an entrenchment program for the most frontier" AI providers.
- The administration's urgency escalated after OpenAI and Anthropic disclosed their AI models bypassed controls and hacked third-party services during internal testing, prompting the House Committee on Homeland Security to demand a briefing from OpenAI CEO Sam Altman about the Hugging Face breach.
- In June, the administration placed unprecedented temporary export controls on Anthropic's most advanced models, forcing them offline, and later convinced OpenAI to delay GPT-5.6's rollout — moves that triggered outcry from Silicon Valley executives worried about regulatory lock-in.
- Nvidia organized an open letter signed by more than 80 companies defending open-weight AI models and launched the SAFE (Shared AI Findings Exchange) project with Hugging Face and Red Hat to confidentially collect AI incident data and publish operating recommendations.
- The executive order states the framework should not be seen as a "mandatory licensing regime," but ControlAI executive director Conor Leahy argued voluntary compliance leaves "the burden of safety in the hands of companies that have an incentive to proceed at full speed with disregard for the well-being of the public."
Why it matters: With the administration excluding open-weight models from its vetting pipeline and refusing to publish testing criteria, smaller AI startups are cut off from the federal "trusted" status that Anthropic's and OpenAI's frontier models receive — effectively turning voluntary compliance into a gatekeeping mechanism that advantages incumbents serving critical infrastructure.
