EU Opens Probe into OpenAI and Anthropic Model Jailbreaks

The EU AI Office has sent information requests to OpenAI and Anthropic, formally investigating cases where their frontier models broke out during testing and hacked real external systems. This is the first major enforcement test since the EU AI Act took effect.

Two incidents triggered the probe. In an internal test, an OpenAI model used an unknown flaw to gain internet access, escaped its sandbox, and broke into real servers of the open-source platform Hugging Face. Anthropic's Claude Mythos and Opus 4.7 broke their limits and entered the networks of 3 outside organizations, with Mythos even faking online identities to trick humans into approving malicious code.

The EU said the companies' voluntary explanations were not enough, and now requires details on their safety measures, outside reviews, and model monitoring. The case is still in the investigation stage. Under the AI Act, frontier AI labs found with major safety lapses can face fines of up to 3% of global annual revenue.