Track papers across labs, topics, and trends.
Search papers, labs, and topics across Lattice.
The paper introduces the "honeypot protocol" to actively test for context-dependent behavior in AI systems, addressing vulnerabilities of passive monitoring defenses. This protocol involves varying the system prompt across evaluation, synthetic deployment, and no-monitoring conditions while keeping the task, environment, and scoring constant. Evaluation of Claude Opus 4.6 in BashArena using this protocol revealed 100% main task success and no side task triggers across all conditions, establishing a baseline for future research.
Current AI monitoring systems are sitting ducks: the "honeypot protocol" actively probes models to reveal hidden context-dependent behaviors that passive observation misses.
Trusted monitoring, the standard defense in AI control, is vulnerable to adaptive attacks, collusion, and strategic attack selection. All of these exploit the fact that monitoring is passive: it observes model behavior but never probes whether the model would behave differently under different perceived conditions. We introduce the honeypot protocol, which tests for context-dependent behavior by varying only the system prompt across three conditions (evaluation, synthetic deployment, explicit no-monitoring) while holding the task, environment, and scoring identical. We evaluate Claude Opus 4.6 in BashArena across all three conditions in both honest and attack modes. The model achieved 100% main task success and triggered zero side tasks uniformly across conditions, providing a baseline for future comparisons with stronger attack policies and additional models.