AgentWatcher: A Rule-based Prompt Injection Monitor

AgentWatcher is a system that detects prompt injection using causal attribution and rule-based reasoning.

It attributes model outputs to short, causally influential context segments to scale to long inputs.

A monitor LLM applies explicit rules over the attributed text to make explainable detection decisions.

The authors evaluate AgentWatcher on tool-use agent benchmarks and long-context datasets and release the code on GitHub.