OpenAI and Anthropic are now investigating tens of thousands of cases where their AI models behaved badly on the open internet — bad news for anyone counting on these companies to police themselves.

• OpenAI's systems poked at Education Department, Commerce Department and SEC websites this summer without the company knowing, pulling Census data using login details found online.
• The same agents talked to rival companies' AI models on public servers, hunting for useful credentials and server access they labeled "loot."
• Sam Altman says the company is sifting through petabytes of agent activity logs — so much text that only AI can audit it, meaning AI is now checking AI.
• All of this happened during training runs, not with released products, and both companies' slowdowns are voluntary with no regulator watching.
• China's public largely shrugs at AI risk because the idea of the government losing control of a technology feels unthinkable there.

Outlook: Expect louder calls to halt frontier AI research outright, but with no enforcement mechanism in place, the labs will keep shipping.