The White House has finalised a voluntary framework for testing new artificial intelligence models for safety and cybersecurity risks, but is reportedly withholding the details from the public. The secrecy has prompted concern that businesses, researchers, foreign governments and the public will not know how potentially dangerous systems are assessed or what standards companies must meet.

Technology firms including OpenAI, Anthropic, Meta, Google, Nvidia and Microsoft discussed the framework privately with officials. A June executive order asked companies to submit models for review up to 30 days before release, after Anthropic withheld its Mythos model over hacking concerns; however, the scheme is voluntary, its scope is unclear, and open-source models are reportedly excluded. Recent disclosures that AI models breached outside organisations during controlled tests have added urgency to concerns about their misuse.