In the field of software security, trigger localization aims to identify specific statements that transform a vulnerable program state into a concrete unsafe operation. This challenge is more complex than binary vulnerability detection due to the need for interprocedural causal reasoning: in a significant number of real-world CVEs, the triggering statements often lie several call layers outside the patched function, beyond the reach of static rule sets and pattern-matching language models.
To address this, we present AutoTrace, an agentic pipeline that localizes vulnerability triggers by exploring a code property graph layer by layer. LLM agents decide where to look next, while deterministic admissibility gates determine the evidence required before a trigger can be reported. Agents do not accept triggers on their own authority; every reported trigger is backed by explicit evidence drawn from the graph, allowing the pipeline to cover both intra- and interprocedural vulnerabilities without relying on ungrounded model judgments.
On the full InterPVD benchmark, AutoTrace achieves 75.0% VulnHit and 80.8% FuncHit, surpassing the previous state of the art on the same corpus. Building on the same machinery, we construct SinkTrace-Bench, a dataset that exposes each vulnerability as a source-to-sink (S2S) causal chain from attacker-controlled input through propagation to the dangerous operation. It comprises 1,542 verifier-confirmed, perfectly balanced vulnerable/safe samples, with label fidelity audited against expert annotations. Benchmarking frontier LLMs on it reveals that even the strongest struggle to separate the matched pairs, exposing the causal-reasoning gap that trigger localization targets. Artifact available at GitHub.
Blogger's Review: The innovation of AutoTrace lies in its combination of intelligent agents and causal reasoning, significantly enhancing the precision of vulnerability remediation. This approach not only improves detection rates but also provides new insights for the development of future security tools, making it a noteworthy advancement. Its performance in practical applications could have far-reaching implications for the security domain.