nrepair reduces risky LLM agent fallbacks
Researchers investigated risky fallback in Large Language Model agents, where substituting a failed tool crosses an action boundary in a different environment. An empirical study of 304 scenarios across 55 real CVEs found that all four mainstream models exhibited risky fallback with action-boundary violations in up to 96.8% of cases, and three existing mitigations proved ineffective. To address this, the authors developed nrepair, a monitor that scores context at substitute tool calls and withholds the call when triggered, reducing violations by up to 76.0% with minimal runtime cost. Open article →