SAFETY SENTRY: Context-Aware Human Intervention via EXECUTE-ASK-REFUSE Routing

The paper reframes safety checking for LLM tool calls from binary safe/unsafe classification to per-instance routing among EXECUTE, ASK, and REFUSE. It introduces Safety Sentry, a lightweight guard model whose inference requires a single decoding call. One decoding-time threshold can reposition the same checkpoint for deployments with different risk tolerances without retraining. The abstract reports higher overall accuracy and safety-related recall than a broad set of open-weight and frontier closed-source baselines, while controlling both directional error rates. However, the supplied record contains no benchmark names, numerical results, model size, or experimental details.