Answers are grounded in this paper and cite their sources.
LLM agents read untrusted content and act through tools with real permissions, so a single injected instruction can turn an agent against its user. Most work tries to prevent such attacks; this paper asks what happens after one succeeds — what evidence agent logs preserve, and whether the attack can be reconstructed without trusting a language model to judge the evidence, since the evidence itself can prompt-inject an LLM investigator. We build a deterministic, LLM-free taint tracer that traces each harmful action back to the untrusted input it drew on (data provenance, not proven causation). On an unseen AgentDojo test of 175 runs it reaches recall 0.98 and 1.00 on the two model families the attack compromised (a third, claude-haiku-4.5, resisted every attack, so detection recall is undefined for it); plain taint (v2) over-flags benign tasks, a delegation-aware version (v3) trades recall for precision, and a locality fix (v4) restores recall to 0.98 with every remaining false positive a genuine flow of untrusted data into a dangerous action. Manual labelling showed the benchmark’s own success labels missed 21 of 58 attacker-directed runs, and a silent hijack emailed an attacker without telling the user. On a 16-scenario controlled LangChain evaluation (simulated tools) the same tracer reconstructs indirect-injection attacks the framework’s native logs cannot; iterative hardening (recipient-trust in v5, address canonicalisation in v6) removes the benign false positives this exposed while keeping full recall across all 16 scenarios, leaving one fundamental blind spot that a tamper-evident flight recorder still records. Finally, on identical traces the deterministic tracer is compared against an LLM-as-judge investigator: an adaptive attacker planting notes aimed at the analyst fools the LLM judge at a model-dependent rate (13–17% on two of three models over 90 trials), while the deterministic tracer, using no language model, is fooled on none.
No takes yet. Share an insight, caveat, or question.
Ahad Nawaz (2026) studied this question.
Synapse has enriched 8 closely related papers on similar clinical questions. Consider them for comparative context: