Skip to main content
Search

# attack-chain reconstruction

1 related articles

Subscribe to updates
  1. arXiv

    AgentTracer traces indirect prompt injection as intent drift, reaching 94.17% injection-point accuracy

    The work presents AgentTracer, an intent-aware tracing framework that treats indirect prompt injection as task intent drift: it recovers implicit decision dependencies among tool calls to build an Intent-Driven Execution Graph, combines the user request with an operation knowledge base to construct a user intent authorization space, and traces backward from an anomalous tool call to reconstruct the attack chain. On AgentDyn and InjecAgent logs combined with a historical-log pool of 1,800 normal requests and 9,000 background tool calls without IPI, it achieves 94.17% injection-point accuracy and 93.56% path precision, improving injection-point accuracy over existing methods by 18–54%.

You’re all caught up