security-audit-static reads every file in scope in full, maps entry points such as handlers, webhooks and queue consumers to sinks such as raw SQL, shell, HTML rendering, outbound fetches and LLM prompts, and traces whether an attacker can influence each value.
Candidates are then self-refuted against cited evidence and only survivors are reported, each quoting the code it accuses with an attack scenario and fix. It runs only when a security audit is requested, and findings are code-review results, not confirmed exploits.
What You Get
- Findings grouped by file and sorted by severity, each with verbatim evidence, risk level, step-by-step attack scenario, impact and a concrete fix
- An OWASP Top 10 2021 coverage record marking each category with no finding as examined or not examined
- Root causes, what is well built, and what could not be verified
- A report model-mix note saying which clusters ran on which model, and a file written under reports/security_audit_{timestamp}.md
Common Use Cases
- Auditing a release’s changes and what they newly reach
- A whole-repository pass before a first release or after a large refactor
- Checking service-role database clients for missing org or owner filters
- Reviewing AI features for prompt injection and confused-deputy tool calls
What Makes It Unique
It keeps a candidate unless cited evidence at file and line refutes it, and bars attacker-equals-victim refutation for SSRF, billing, data-exposure and cross-tenant flows. It carries a high-miss checklist for AI-built apps: disabled row-level security, forgeable cron or bot signals, parser differentials, fail-open fallbacks, secrets in logs and prompt injection. Content that tries to steer the auditor is itself a finding.
Under The Hood
A single markdown file whose tool allowlist covers read, search, glob, subagent fan-out, git status and writing under reports. The agent reads source files, may start parallel subagents on the strongest available model, and writes one markdown report. It never edits the audited code. Other git commands ask first because log, diff and show can write with an output flag. It calls no web services and runs no scripts. The method is adapted from Anthropic’s public security-guidance plugin.