Evidence · DiggingBeagle record
AgentTracer: tracing indirect prompt injection through intention-execution alignment
Original study of reconstructing injection paths from reasoning and tool-call histories on two controlled benchmarks.
- Published
- Oct 7, 2026
- Source role
- primary disclosure
Evidence record
Original study of reconstructing injection paths from reasoning and tool-call histories on two controlled benchmarks.
Claim-level citations (3)
- supportsAgentTracer reconstructs indirect prompt-injection paths from agent tool histories: AgentTracer reported 97.34% detection, 94.17% injection-point accuracy and 90.25% path coverage across its two selected 300-trace benchmark sets.
Section 5.2, Table 2: End-to-end Tracing Effectiveness
- supportsAgentTracer reconstructs indirect prompt-injection paths from agent tool histories: The paper identifies incorrect LLM graph relation extraction and entity alias grounding as causes of tracing errors.
Section 5.2, Reasons for tracing errors
- supportsAgentTracer reconstructs indirect prompt-injection paths from agent tool histories: The benign-task evaluation reported an edge-level 3.52% false positive rate, 13 false-positive edges among 369 legitimate edges after separately reviewing genuine task deviations.
Section 5.5, Result Analysis: 25 genuinely deviating edges, 13 false positives of 369 remaining legitimate edges