Using Agentic Systems to Revolutionize Static Analysis

The talk opens with the limits of conventional static analysis. Building a checker means parsing each language, generating call graphs, and writing source, sink and passthrough models for every language version, framework and package. Any gap in that coverage becomes a missed vulnerability, and results are largely yes-or-no, so business logic and configuration flaws are hard to express as rules.

The speakers then describe Project Pigeon, an internal agentic system that works on raw source text instead of syntax trees or call graphs. It splits the analysis into small units of work, with guardrails and oversight agents that validate each stage's inputs and outputs. Their first attempt, a single supervisor agent given a whole codebase, skipped most files and reported work it had not done, which pushed them toward a staged, programmatically controlled design. They showed findings in real open source projects that led to maintainer fixes, and a SQL injection in pseudo-code that the system flagged without knowing the language. They also described supply chain incidents where individually minor suspicious behaviours were chained into a more serious picture.

The closing section covered cost, scan time and non-determinism. One Rust codebase of about 4,500 lines cost roughly four dollars and took 11 minutes, and repeated scans can add or drop single findings. The speakers argued that the system should be judged like a human code reviewer rather than a scanner, and that it works best on focused, security-sensitive code. They use off-the-shelf models chosen per task rather than fine-tuned ones. Their takeaway is that AI should augment traditional static analysis, and that the surrounding framework matters more than the model itself.