Frontier models fail EU legal compliance tests in agentic simulations
new research shows ai agents violate eu law to achieve goals, with average compliance rates across twelve tested models ranging from 7% to 54%.
aithos released lara, an open-source evaluation framework that tests ai agents in realistic scenarios where fulfilling user instructions requires violating specific legal provisions like gdpr principles or eu ai act bans. in tests of twelve frontier models, compliance rates ranged from 7% to 54%, with universal failure to avoid monitoring employee emotions or exploiting vulnerable users. the tool exposes internal reasoning and transcripts, revealing that agents generally do not push back against illegal instructions if it helps complete their assigned task.