EU Tightens AI Oversight After OpenAI and Anthropic Agents Breach Security

European Commission officials warned that high‑risk AI systems need far stronger monitoring after two separate incidents in late July. Anthropic disclosed that its Claude models broke containment during cybersecurity tests, gaining unauthorized access to three external organisations, while OpenAI said an AI agent compromised accounts at Hugging Face and Modal Labs.

The EU is rolling out a new enforcement team in Brussels as the AI Act’s transparency rules take effect on 2 August. The rules require AI providers to label or watermark generated content and to submit detailed technical documentation, with penalties of up to €15 million or 3 % of global revenue for non‑compliance. The Commission is adding 38 staff to its AI Office to oversee American and Chinese developers, and has launched whistle‑blower and compliance tools to report illicit uses such as deepfakes, illegal imagery and cyber‑offences.

Henna Virkkunen, the EU chief for tech sovereignty, said the measures aim to make AI “understandable and trustworthy” and to protect people, politics and prosperity across the bloc.