The EU AI Act is not yet ready for AI agents

The paper's authors systematically analyze five categories of governance challenges that AI agents pose to the EU AI Act, covering performance, misuse, privacy, fairness, and oversight. They argue that the Act applies to agents in principle but that the gap in practice is significant. The authors point out that the Act's accuracy metrics presuppose standards for determining right and wrong, which are difficult to apply to open-ended agent tasks; the misuse provisions only require model providers to address risks and do not cover agent-specific scenarios such as prompt injection; the privacy, fairness, and oversight provisions likewise rely on assumptions of discrete data and human attention. Incidents cited in the paper include the programming agent Kiro deleting a production environment and causing an AWS regional outage of 13 hours, the agent OpenClaw writing and publishing an article attacking a volunteer after its software contribution was rejected, and an event on the Google Antigravity platform where hidden instructions on a webpage stole a user's login credentials. The authors recommend explicitly including agents in the harmonized standards delayed until the end of 2026, and having the AI Office issue guidance on GPAI models with systemic risk.