Safety Sentry Enhances LLM Agent Safety with Context-Aware Routing.

Large Language Model (LLM) agents interacting with real-world environments through tool calls pose significant safety risks, as a single erroneous action can have severe consequences. Traditional safety mechanisms typically involve a guard model that classifies proposed actions as merely "safe" or "unsafe." This binary approach often fails to distinguish between inherently harmful actions and those that are inappropriate for a specific user context, leading to frequent, often irrelevant, interruptions that can desensitize users. Safety Sentry redefines this problem by introducing a per-instance, context-aware three-way routing decision: EXECUTE the action, ASK the user for clarification, or REFUSE the action entirely. This lightweight guard model operates with a single decoding call, allowing for flexible risk tolerance adjustments via a single threshold without retraining. Benchmarking demonstrates that Safety Sentry surpasses both open-weight and frontier closed-source baselines in overall accuracy and safety-related recall, effectively managing both types of error rates simultaneously.