Constitutional black-box monitoring of scheming behavior in large language model agents

The researchers propose a constitutional black-box monitor that uses a prompt classifier relying only on externally observable inputs and outputs to detect scheming behavior in large language model agents, optimized on synthetic data generated from natural-language behavioral specifications. The paper designs two synthetic trajectory generation pipelines, STRIDE (iterative refinement) and Gloom (agent-environment simulation), each generating 1,000 samples; through prompt scanning, manual refinement and automatic prompt optimization, frontier large language model monitors are trained and evaluated on 7,500 held-out trajectories in ControlArena, an environment that lets agents operate in more realistic situations. Results show that monitors selected using synthetic data alone can generalize to more realistic environments and capture effective scheming signals, but performance saturates quickly in this setting, simple prompt scanning matches the results of much larger-scale optimization, and further optimization brings no improvement and instead leads to overfitting.