Broken Actors Jun 2026

| Approach | Method | Limitation | |----------|--------|-------------| | Behavioral fingerprinting | Compare action distributions to a baseline | Broken actors may mimic normal behavior partially | | Reward auditing | Monitor reward components for implausible strategies | Requires interpretable reward structure | | Cross-agent consensus | Require multiple agents to agree before acting | Slows down system; vulnerable to collusion | | Sandboxed rollouts | Periodically test agents in simulated environments | Computationally expensive |

Limited compute or memory causes erratic behavior. Example: A drone with low battery recalculates its path every second, oscillating wildly. It is still “running” but broken in practice. broken actors