Researchers have a new way to pinpoint exactly which log entries pushed an AI security assistant toward its conclusion.
The team proposes Topological Attribution Distance (TAD), a technique for tracing which segments of retrieved evidence, such as cyber incident logs, most shaped an LLM's output in retrieval-augmented generation systems. Rather than just noting which documents were retrieved, TAD measures how much a model's response shifts geometrically in embedding space when a specific log segment is removed - a method the authors call segment-level ablation attribution. They tested the approach on logs from an actual cyberattack and showed it can adaptively surface the log entries that mattered most to a given output. The goal is a decision trail: when an autonomous security agent flags something, defenders can trace that call back to the specific evidence behind it.
That matters because security teams are increasingly letting agentic AI handle triage and response, and auditing those calls is becoming a trust and compliance requirement, not a nice-to-have. Standard RAG citation tools tend to show what was retrieved, not which retrieved passage actually swayed the answer - a gap that shows up fast when incident logs are near-duplicates and easy to conflate.
This is a research proof of concept tested on one incident, not a shipped product. Whether "geometric shift" attribution holds up on messier, larger log sets - and whether security teams trust topology math more than the black box it's replacing - is still an open question.