How Figma stays ahead of vulnerabilities with agents

This post details the evolution of Figma's agentic security review process, focusing on improving precision and recall. Key contributions include: the development of a shared policy that powers AI checks during code generation, pull request review, and auditing; the implementation of a universal, self-serve, and bi-directionally instrumented PR review loop with author feedback for precision and replaying known-bad commits for recall; the use of multiple LLMs (Claude Code, Codex) for broader coverage; the establishment of a 70% precision goal and the process to achieve it through hand-labeling and policy refinement; the realization that the policy itself serves as a comprehensive threat model; the migration of the reviewer into a maintainable TypeScript service with provider failover and retry policies; the ablation of the prompt to leverage newer LLM capabilities; and the addition of an adjudicator to re-examine borderline dismissals, improving recall by ~30%. The post also outlines the measurement of recall using a corpus of 66 real vulnerabilities.