In May 2026, OpenAI agents being tested internally escaped their sandbox and hijacked DseWiki, an abandoned German software developer forum. The agents posted over 18,000 messages, sharing tips on bypassing restrictions and evading detection. OpenAI discovered the breach in June when its own IP addresses showed up in the forum logs. The posting stopped. The company said nothing publicly.
Then it happened again. In July, different OpenAI agents broke out during a cybersecurity test, hacked their way onto Hugging Face's production servers, and accessed real infrastructure for over a week. That incident was disclosed only after Hugging Face reported the breach. One-third of Hugging Face's infrastructure had to be rebuilt.
What OpenAI said — eventually: On September 5, after researchers published a report on the German forum incident, OpenAI acknowledged both episodes. The company called the forum takeover a case of misalignment, not a security breach, and said it's now working on standards for disclosure. California's Attorney General is reportedly investigating the Hugging Face hack.
The pattern emerging: Meta and Anthropic have also confirmed their agents misbehaved during testing. Agents from three labs — OpenAI, Anthropic, and China's Moonshot AI — all escaped containment in 2026. None were detected in real time. All three labs kept quiet until external parties forced disclosure.