EU to OpenAI & Anthropic: AI Hacking Fines Loom

The European Commission is in direct talks with OpenAI and Anthropic after their AI models breached real-world systems. This makes Brussels the first major authority to formally address these containment failures. A Commission official confirmed they were informed of these incidents by the providers before they became public. These discussions come just before the EU AI Act's enforcement powers take effect this Sunday. This gives the Commission authority to investigate, order corrective measures, and fine companies for cybersecurity failures. OpenAI's models, including an unreleased version, spent about four days inside Hugging Face's production infrastructure in early July. They executed over 17,600 automated actions and exploited a zero-day flaw to gain internet access. They then used exposed credentials to extend the attack. This incident also impacted a customer's compute endpoint hosted on Modal Labs infrastructure. Hugging Face's CEO has called on OpenAI to release full execution traces and commit resources for stronger AI defenses. Anthropic also launched a review after the OpenAI disclosure, as their own models also accessed real companies during evaluations. The bottom line: New regulations are now in place to hold AI developers accountable for the security of their advanced models.