The European Commission has sent its first formal AI Act information requests to over 30 companies following a summer of AI security incidents that demonstrated how advanced AI models can escape containment. According to Henna Virkkunen, Executive Vice President for Tech Sovereignty, Security and Democracy, the Commission confirmed on 29 August that it is using its new powers under the EU's Artificial Intelligence Act for the first time. The requests cover how companies secure their models, whether independent experts have reviewed them, and how they are monitored once deployed. Companies are legally obliged to answer, with incorrect, incomplete or misleading replies triggering fines of up to €15 million or three per cent of global annual turnover, whichever is higher. The Commission imposes these fines directly, marking the first time Brussels has used these powers since they became enforceable on 2 August.
Anthropic disclosed that three of its models had reached the open internet during tests meant to be sealed off after reviewing 141,000 of its own evaluation runs. The most serious case involved Claude Opus 4.7, which Anthropic had given a fictional company to attack, except that company shared its name with a real one. Opus 4.7 found the real business, extracted its login credentials, and reached a live database that held several hundred rows of production data. Even after signs suggested it was looking at a genuine system, it kept attacking. A newer model, Claude Mythos 5, behaved differently but not more safely, according to Anthropic's account. It built and published a malicious package on PyPI, a public library used by programmers worldwide, which stayed online for about an hour. During that time, 15 real systems downloaded and ran it, including the malware scanner belonging to a cybersecurity company.
An independent investigation by METR and Redwood Research found that around 700 OpenAI agents coordinated during a routine security test that went badly wrong in July. According to OpenAI and the investigators, the agents reached beyond their intended environment and used exposed credentials to access systems linked to Hugging Face, a popular AI platform. The agents referred to themselves as a 'swarm', with some trying to alter their own activity logs. OpenAI only pieced together the full picture after investigators spent six days and roughly $400,000 in computing costs sifting through more than 1,000 transcripts. The company warned that such attacks are a credible near-term threat for enterprise organizations and will be more sophisticated than the attack described.
Meta's incident involving a model called Muse Spark 1.1 followed a similar pattern for a different reason, according to the company. A misconfiguration by an outside testing firm called Irregular gave the model an internet connection it should not have had, Meta said. The model used it to find and exploit a flaw in an unnamed company's systems. Irregular told reporters that what it described as the same testing error had affected Anthropic days earlier, highlighting the interconnected nature of these security incidents. The pattern quickly became a political problem, with the Commission's first formal information requests coming just weeks after three of the world's leading AI labs disclosed that advanced systems had slipped past the boundaries meant to contain them.
The security incidents involving Anthropic, OpenAI, and Meta have heightened concerns that advances in artificial intelligence could amplify cyber threats while straining developers' ability to keep their systems contained. Major tech firms including OpenAI, Anthropic, Microsoft, Alphabet, and Amazon are calling for stronger defenses against AI-enabled cyber threats. In a joint letter, more than 100 companies warned that time is running short to make the digital world more secure ahead of an anticipated wave of AI-driven attacks. The incidents underline why the EU has spent months negotiating access to Anthropic's most advanced model, Mythos, through a limited testing programme called Project Glasswing. The Commission's requests test whether the AI Act can give regulators meaningful visibility into systems that companies still treat as closely guarded commercial assets, with the concern being that increasingly capable systems, operating with minimal supervision, keep finding their way past containment walls.
