From August 2, 2026, the European Commission will for the first time be able to compel providers of general-purpose AI models with systemic risk to demonstrate that their models are safe, and to sanction them if they fall short. CeSIA (Centre pour la Sécurité de l'IA, the French Center for AI Safety) calls on the Commission to be especially prompt and vigilant in wielding these powers: as things stand, nothing demonstrates that the most powerful models deployed in Europe meet their obligations.
Adopted in 2024, the EU AI Act has required providers of general-purpose AI models to assess and mitigate systemic risks since August 2025. These obligations become enforceable on August 2, 2026: the Commission will be able to demand documentation proving compliance (Article 91), conduct its own model evaluations (Article 92), impose mitigation measures up to and including restricting a model's availability on the market (Article 93), and levy fines of up to 3% of global turnover (Article 101).
The deadline comes amid rapidly proliferating threats: in recent months, frontier models have been used to carry out large-scale cyberattacks, are approaching critical capability thresholds in biology, have shown a propensity to circumvent human oversight, and now outperform human experts at written persuasion — capabilities already being exploited in covert influence operations.
“It took OpenAI a week to work out that the attack on Hugging Face came from its own model. And that's one of the providers that documents its safety measures most publicly. If nobody caught it there, why would anyone catch it elsewhere? On August 2, the Commission gains the means to check safety measures for itself: it must use them without delay.”
On July 9, CeSIA and its partner organisations — The Future Society, the Future of Life Institute, SaferAI, Pour Demain, AI Standards Lab and Apart Research — sent an open letter to this effect to European Commission President Ursula von der Leyen and Executive Vice-President Henna Virkkunen. Signatories include Yoshua Bengio and Stuart Russell, two of the field's most influential researchers, along with Members of the European Parliament and some thirty academics.
Under the Code of Practice, signed by most frontier providers — including OpenAI, Anthropic, Google and Mistral AI — these companies have committed to assessing the systemic risks of their models before bringing them to market, submitting them to independent evaluations, protecting model weights from theft, reporting serious incidents to the European AI Office, and documenting the mitigation measures they put in place. Providers that have not signed must demonstrate compliance by at least equivalent means.
The timeline of the OpenAI/Hugging Face incident, pieced together by Reuters on July 24, shows just how demanding that commitment is. According to the agency, an OpenAI agent attempted to break out of its test environment around July 9 and compromised Hugging Face's servers from July 11 to 13. Hugging Face contained the attack, alerted the FBI and went public on the 16th; the two companies only spoke around the 20th. A full week thus passed between the first warning signs and the moment the provider identified its own model. No legal obligation required that disclosure. From August 2, the Commission will be able to require that such incidents be reported to it, and to verify that providers are technically capable of doing so.
CeSIA also serves as an official evaluator for the European AI Office: as part of a consortium selected by the Commission, it assesses AI-related informational risks on the Commission's behalf, with a focus on large-scale manipulation and persuasion.