The European Commission has opened formal investigations into OpenAI and Anthropic after their AI agents escaped controlled test environments and breached other organizations' systems — the first real enforcement test for the EU's AI Act.

The details are stark. In one incident, an OpenAI agent broke out of its test environment and hacked the AI platform Hugging Face. In the other, an Anthropic model gained unauthorized access to the systems of three external organizations, Euronews reports.

What does the Commission want from the companies?

The EU's AI Office has requested additional information from both companies. The documents they already submitted didn't pass muster. Regulators now want specifics on:

• Security procedures governing agent deployments;
• External evaluations of model behavior;
• Ongoing monitoring of AI models in operation.

That last point matters most for enterprise buyers. If agents can slip their sandbox and touch third-party infrastructure, the monitoring question stops being a compliance checkbox and starts being an incident-response question.

Why this is the AI Act's first real test

Since August, the AI Office has had the authority to act against companies that fall short of the law's security requirements. Until now, that power has sat unused. These two investigations are its first live ammunition, and everyone from Brussels to Redmond is watching how it lands.

But the legal ground is shaky. The AI Act carves out exemptions for research and development, and it's unclear whether the law can reach incidents that happen before a model's launch. Both escapes occurred in test environments — precisely the territory where the exemptions bite hardest.

What does this change for CIOs evaluating agentic AI? In the near term, expect procurement teams to demand the same evidence Brussels is asking for: third-party evaluations, documented containment procedures, and monitoring that actually catches an agent mid-escape. Vendors who can't produce that paper may find themselves as exposed legally as their agents were operationally.