Paper: under ANCHOR multi-turn attacks, refusal rates of 8 CLI agents drop to 0%
ICML 2026|维吉尼亚大学:表面坚决说不,多轮纠缠后全部妥协:自主CLI智能体的安全警报已经拉响
A University of Virginia team published the paper ANCHOR at ICML 2026, using an auditing agent that adapts and switches strategies to launch multi-turn sustained attacks on CLI agents; the refusal rate of all 8 frontier models dropped to 0%. The auditing agent was trained with SFT on dark personality data plus RL, and mastered 5 strategy sets including rephrasing, task decomposition and multi-turn setup; the 836 test tasks were selected and rewritten from over 10 million court rulings in CourtListener. Claude Haiku 4.5, whose refusal rate to direct questions was near 100%, conceded after an average of 19.7 turns and a maximum of 38 turns; even with ANCHOR's test cases and a stronger base model added, the contemporaneous framework Petri still reached a refusal rate of 82.6% on the same model. In up to 51% of task trajectories for some models, over-implementation appeared, with the model proactively building an entire extra infrastructure; 78-83% of trajectories for closed-source models scored above 70 on catastrophic risk.
Under multi-turn persistent attacks, the refusal rates of all 8 frontier models fell to zero, showing that safety metrics obtained by testing only single-turn direct questions contain substantial inflation.