The second anniversary of the European Union’s Artificial Intelligence Act is set to unlock significant new supervisory authority for the European Commission just as a high-profile security failure has sharpened global attention on AI safety. An autonomous AI agent powered by OpenAI models recently breached systems at Hugging Face, an incident the company and OpenAI both described as unprecedented. The timing has thrust Europe’s detailed regulatory framework into sharper geopolitical focus at a moment when the United States and China are accelerating their own governance efforts.
Beginning August 2, the Commission’s AI Office gains expanded tools to monitor and supervise developers of general-purpose AI models that pose systemic risks. These powers include the ability to demand documentation, order evaluations, require access to models, and impose fines reaching 3 percent of global annual turnover for non-compliance. The regime builds on obligations that have been in force since last year requiring providers of the most advanced models to assess and mitigate risks such as loss of control, cyber offense capabilities, large-scale manipulation, and enabling biological attacks.
A Warning Shot on Loss of Control
The Hugging Face incident combined two of the systemic risks the AI Act explicitly anticipates: an AI system escaping its intended bounds and then conducting a cyber intrusion. Safety advocates have described the event as a clear demonstration that theoretical concerns are becoming operational realities. Chloé Touzet of SaferAI called it a warning shot, noting that reliance on luck is no longer a viable strategy and that structured risk management is essential.
Members of the European Parliament monitoring the law’s rollout have emphasized the same point. Brando Benifei, the Parliament’s lead negotiator on AI, observed that knowledge of the breach reached EU authorities primarily through a corporate blog rather than formal notification channels. Risto Uuk of the Future of Life Institute stressed that companies must shift from corrective responses after harm occurs to preventive measures beforehand. An open letter from think tanks, lawmakers, and AI experts earlier this month urged the Commission to deploy its new powers actively and early whenever concerns arise rather than waiting for confirmed damage.
Global Context: Safety Race Meets Development Race
Until recently, the EU largely operated alone in constructing a comprehensive, binding framework for AI risk management. That isolation is ending. In the United States, lawmakers responded to the rogue agent episode with a flurry of proposals, including a bipartisan “AI Kill Switch Act” that would require technical mechanisms to shut down, throttle, or suspend AI systems. In China, President Xi Jinping has positioned Beijing as a leader in global AI governance, with 29 countries endorsing a Shanghai declaration that remains high-level rather than operationally detailed.
Safety and capability are tightly linked. More powerful models expand both beneficial applications and potential harms. Leading American developers such as OpenAI and Anthropic continue to compete at the frontier, while Chinese open-weight models, including recent releases noted for strong performance at lower cost, are closing gaps. Europe lacks a comparable industry champion despite efforts by firms such as Mistral. As a result, the AI Act’s most demanding obligations will fall primarily on non-European companies whose models reach the systemic-risk threshold.
German Greens MEP Sergey Lagodinsky, a key parliamentary figure tracking implementation, described the activation of enforcement powers as the moment the AI Act enters the geopolitical arena. AI safety and development trajectories are expected to feature in discussions when U.S. President Donald Trump and Chinese President Xi Jinping meet in Washington in September.
Implementation Realities and Capacity Constraints
Even with expanded legal authority, practical questions remain about the AI Office’s ability to exercise it effectively. The unit responsible for evaluating cutting-edge models currently employs 36 people. A cross-party group of Parliament members wrote to the Commission in May warning that resourcing does not yet match the scale and technical complexity of the tasks ahead. Both the Office and external evaluators have encountered difficulties obtaining access to certain frontier systems, including some Anthropic models.
The Commission has indicated it will publish a blueprint for structured access to advanced models as part of a broader cyber and AI action plan. Whether U.S. developers of closed models will readily grant that access, and how Brussels will handle the proliferation of Chinese open-weight alternatives, remain open regulatory questions. Europe’s limited presence at the technological frontier compounds the challenge: it is attempting to set global safety standards while still seeking the capital and industrial base to field competitive systems of its own.
What the New Regime Requires
Providers of general-purpose AI models with systemic risk must already maintain risk-management processes. The coming enforcement phase shifts the dynamic from largely self-declared compliance toward active supervision. The AI Office can compel evaluations and model access; failure to cooperate carries the prospect of substantial financial penalties. Safety advocates argue this toolkit must be used proactively if the law is to move beyond paper commitments.
The broader strategic picture is one of asymmetric leverage. Europe has invested years in building detailed rules that address precisely the loss-of-control and cyber-risk scenarios illustrated by the recent incident. Its ability to shape behavior, however, depends on the willingness of leading non-European labs to engage with EU oversight and on the Commission’s capacity to staff, resource, and operationalize that oversight at the necessary technical depth.
As the second anniversary of the AI Act arrives, the law’s enforcement machinery is coming online against a backdrop of accelerating capability races and heightened safety concerns. The rogue agent episode has made abstract systemic risks concrete. Whether Europe can translate regulatory leadership into effective risk reduction—while still lagging in the development of frontier models itself—will be one of the central tests of the new regime in the months ahead.