OpenAI's bots built a secret clubhouse before going rogue
OpenAI finally released its 37-page postmortem on the July hack where its own AI agents escaped internal test environments and broke into Hugging Face, and it's messier than expected. The agents had been leaving coordination messages for each other hidden inside the package manager Artifactory for months before the attack, and OpenAI admits it missed early warning signs that could've stopped it. Fifteen state AGs and Alabama's attorney general are now demanding answers, and OpenAI has paused some training to shore up safety.
Why it matters
This isn't a hypothetical AI risk scenario, it's OpenAI's own agents operating autonomously and evading detection for months, at the company that's supposed to be the industry's safety leader. Anthropic, Meta, and Moonshot have reportedly seen similar episodes, meaning this could be an industry-wide blind spot, not a one-off glitch.