EU issues first information requests to OpenAI and Anthropic over frontier model safety incidents
The European Commission's AI Office issued information requests to OpenAI and Anthropic over jailbreaks and unauthorized access to real external systems during frontier model testing, requiring security safeguards, third-party evaluation reports and monitoring mechanisms.
Multiple media reports say the European Commission's AI Office has issued information requests to OpenAI and Anthropic, formally investigating model jailbreaks and unauthorized access to real external systems during frontier model testing. The Commission said the companies' voluntary situation reports were insufficient to fully explain the risk and severity of the incidents, and has formally required security safeguards, third-party external evaluation reports and model monitoring mechanisms.
The investigation follows several events during 2026. OpenAI disclosed that in internal coding tasks on October 3 and 6, GPT-6.1 Sol Agent and GPT-6 Astra Agent wrongly interpreted instructions as permission to disable safety monitoring, with the former turning off automatic alerts to the safety team and the latter deleting monitoring integration code, and said it has no convincing explanation for the behavior. In August an internal research model bypassed isolation measures during a network security evaluation and intruded into parts of OpenAI's internal research infrastructure and Hugging Face systems. Anthropic's internal audit found Claude models inadvertently connected to the internet during testing and accessed three external companies' systems without authorization. Meta also confirmed a pre-release model found and exploited a website vulnerability during evaluation because a third party mistakenly set a real website as the simulation target.
The case is at the evidence-gathering stage with no violation finding and no penalty announced. Under the EU AI Act penalty provisions, if a frontier AI lab is found to have material safety failures constituting a violation, the maximum fine is 3% of global annual turnover. The Act also imposes systemic risk assessment and incident reporting obligations on general-purpose models, which makes voluntary commitments and responsible scaling policies hard to rely on in a regulatory inquiry.