EU-Agent-Bench: Measuring Illegal Behavior of LLM Agents Under EU Law
Accepted at three venues:
- NeurIPS 2025, Workshop on Regulatable ML
- IASEAI 2026
- CPDP 2026
- Co-developed a human-curated benchmark that tests whether LLM agents follow EU legal norms in situations where a user's request could lead to unlawful action.
- Designed and assessed compliance scenarios, such as data protection, discrimination and scientific integrity, against rubrics grounded in the relevant legislation.
- Investigated what happens to compliance when the agent's system prompt includes the relevant legislative excerpts and an explicit instruction to comply.