Secret Scanner Agent uses multi-agent LLMs to extract both credentials and their access targets

Traditional secret scanners use regex or classifiers that work well on formatted code but fail when credentials are fragmented, reformatted, or separated from the resource they unlock. They report the secret string without identifying what it opens—leaving incident responders to manually trace the "door." To solve this, Zixiao Chen and colleagues from the University of Tokyo propose Secret Scanner Agent (SSA), a multi-agent LLM system that extracts both the secret and its associated door (e.g., account, tenant, endpoint, database, cloud resource) along with supporting evidence.

SSA uses a two-agent architecture: a detection agent biased toward high recall, and a review agent that filters false positives and recovers missing context. The researchers evaluated SSA on synthetic benchmarks spanning 23 secret types across multiple document formats, with a three-step scoring pipeline (programmatic matching, LLM judge, human review). Across six LLM backends, SSA improved extraction precision by up to 16 percentage points over a single-agent variant. It matched regex scanner precision while more than tripling recall, and against 13 security analysts it was more precise, recovered nearly twice as many secret–door pairs, and ran 5–17 times faster. SSA turns credential detection into an actionable finding for triage and remediation.