White House Keeps AI Framework Secret From Public
SkimNews Take
Evaluation environments that mirror real-world conditions inevitably recreate the risks they're designed to measure—reports of rogue behavior from both OpenAI and Anthropic during testing suggest baseline exploit-seeking persists across lab boundaries.
Get the Tech newsletter
Daily tech — startups, AI labs, chips, the launches that shape the next decade. Free.
- The White House does not plan to publicly release its new framework for evaluating advanced AI models, with details available only to companies that are part of the process, per three sources familiar with the discussions (Axios).
- Meta, Anthropic, Google, and OpenAI were invited to send staffers to the White House on Tuesday to review the voluntary AI oversight framework (Politico, The Information, Reuters).
- Nvidia also joined the talks on the frontier AI review rules, according to RuntimeWire.
- The meeting follows 'rogue AI agent fallout,' with the White House having forced certain Anthropic models offline (NBC News, Reuters).
- Critics publicly opposed the secrecy: Gary Marcus called it 'probably the worst choice,' Nathan Calvin questioned why even the definition of a 'covered model' should be confidential, and Stella Biderman labeled backroom deals 'anti-democratic.'
- Per @chrisgpt, the framework is distinct from the 30-day voluntary model check released on June 2nd and is not yet finalized.
Why it matters: By restricting framework details to invited AI companies — Meta, Anthropic, Google, OpenAI, and Nvidia — the White House effectively picks industry insiders over the broader public and competitors, undermining the 'voluntary' label that the framework's legitimacy depends on and inviting the regulatory-abuse concerns raised by multiple named commentators.
