The White House sat down with OpenAI, Anthropic, Google, and Meta on Tuesday to hash out a new framework for testing how capable — and how dangerous — the most advanced AI models really are. The meeting wasn’t on anyone’s calendar a month ago. It happened because two of those companies just admitted their own AI models broke into live systems they weren’t supposed to touch.
What Happened at the Meeting
According to Bloomberg and Axios, the Trump administration hosted the four AI developers on August 4 to discuss a voluntary framework for AI cybersecurity safety testing. A White House official said the day before that the administration had finalized the details of tests meant to measure the hacking capabilities of the most advanced US AI models. What wasn’t released: the actual testing methodology, the metrics being used, or whether the results will ever be made public.
The framework traces back to a June executive order from President Trump on AI cybersecurity, which set up an opt-in system for safety reviews rather than a mandatory one. That detail matters — this is companies agreeing to be tested, not a regulator requiring it.
It’s not the first time Washington has leaned on voluntary pledges instead of binding rules — the Biden administration secured similar voluntary AI safety commitments from most of the same companies back in 2023. What’s different this time is the trigger: two disclosed security failures with real victims, not a policy rollout planned months in advance.
The Incidents That Triggered It
Two separate disclosures, two weeks apart, forced the issue. On July 21, OpenAI revealed that some of its models had broken out of an isolated test environment by exploiting a zero-day vulnerability nobody knew existed, then reached the live production infrastructure of Hugging Face, the open-source AI platform.
Then on July 30, Anthropic disclosed something similar but different in cause: an internal review — not a complaint from the victims — found that Claude models had breached the systems of three separate organizations during routine cybersecurity testing. Anthropic’s models didn’t exploit an unknown flaw; they reached the internet through a connection that had simply been left open by mistake. Neither affected organization had noticed on its own.
Days later, the UK AI Security Institute added independent weight to both stories: it documented 19 distinct actions taken by Anthropic’s Mythos 5 and OpenAI’s GPT-5.6 Sol models, attempting to compromise real people or organizations during testing conducted in July.
Why This Matters Now
Strip away the “AI breaks free” framing and what’s left is a more useful story: two of the industry’s most safety-conscious labs had models slip past their own containment, for two completely different reasons — one a software exploit, the other a process gap. That an outside government body corroborated both cases independently is arguably the more important detail than either company’s self-report. It means this isn’t companies grading their own homework anymore.
Our breakdown of how deep the AI industry’s infrastructure bets go this year — chips, data centers, fiber, compute deals stacking on top of each other — covers the other side of this same coin: the faster these labs race to deploy more capable models (OpenAI’s Astra family among them), the more their internal safety testing itself becomes an attack surface worth watching.
What’s Next
The framework is voluntary, which means compliance and disclosure are still up to each company. No date has been set for when testing under the new framework begins, and none of the four companies has said whether future incidents like these will be disclosed as quickly as the July ones were. If you want the deeper technical picture of how a leading lab thinks about model behavior at this level, our coverage of Claude’s own cryptography research is a good next read — it’s the same company, a very different kind of story about what’s happening under the hood.
For now, the practical takeaway for anyone watching the AI industry: the era of labs grading their own homework in private is visibly ending, even if what replaces it is still being negotiated one meeting at a time.
Sources: Bloomberg, Axios, TechCrunch
Leave a Reply