Security audit

security-audit

Adversarial multi-agent security review. Different from pr-review's broad code review — security-audit's three lenses are explicitly InfoSec-shaped, the severity vocab is CVSS-aligned, and the output is a threat model rather than a code-review findings list.

How agents collaborate

Lens	Agent	Looks for
Auth + data flow	claude	Trust boundaries, identity propagation, authorization vs authentication, data flow to sinks
Injection + priv-esc	codex	SQL/command/XSS injection, path traversal, SSRF, race conditions, crypto misuse
Dependency + supply-chain	antigravity	Known CVEs, version drift, transitive trust, license compatibility

Severity vocab (CVSS-aligned):

• critical — RCE, auth bypass, privilege escalation
• high — data breach, integrity violation
• medium — DoS, info disclosure, weak boundaries
• low — defense-in-depth gap, timing leak
• informational — worth noting but no exploitable path

Two input modes

# Diff mode — audit a PR
jutsu swarm security-audit --pr 42

# Diff mode — audit a local branch (no PR yet)
jutsu swarm security-audit --diff-from-branch origin/main

# Files mode — audit specific paths
jutsu swarm security-audit cmd/server/auth.go

# Files mode — audit a directory (multi-file concatenated)
jutsu swarm security-audit cmd/server/

# --quick to override the default --full mode
jutsu swarm security-audit --pr 42 --mode quick

# Iterate cheap on synth prompt
jutsu swarm security-audit --replay <key>

The two modes are mutually exclusive — passing --pr AND positional file paths errors with a "pick one" hint.

--full defaults ON

Unlike other presets, security-audit runs --full mode (with Pass-2 round-robin debate) by default. Security findings have asymmetric cost: false-positives erode user trust faster than missed issues. The debate round filters false-positives by having each agent critique its peers.

Override to --mode quick only when you're doing a fast first-pass scan.

First-run consent

Before the first invocation in a repo:

jutsu swarm security-audit --grant-consent

Persists allow-multi-model: true to .kaijutsu/security-audit.yaml. Diffs/files go to remote model providers; explicit consent required.

Important caveat for security-audit specifically: the input may contain code that itself contains secrets, vulnerabilities, or proprietary IP. The pre-flight secrets-scan blocks any obvious credential leaks (.env files, AWS keys, etc.) — but it doesn't catch every form of sensitive data. If the audit target has compliance or NDA constraints, consider running on redacted excerpts or skip security-audit entirely in favor of a self-hosted setup (Phase 3+).

Reading the output

### Threat model

#### Auth bypass via missing token validation — critical (3/3 reviewers)
- **Where**: `cmd/server/auth.go:42-58`
- **Attack vector**: Anonymous request with no Authorization header reaches the protected handler because the middleware short-circuits on err != nil.
- **Impact**: Full account takeover for any user whose ID an attacker can guess.
- **Mitigation**: Treat ErrTokenMissing as 401 explicitly; don't let the err-handling fall through.

#### SQL injection in search query — high (2/3 reviewers)
- ...

#### Stale dependency with active CVE — medium (1/3 reviewers, antigravity)
- ...

### Disagreements
- 1/N rows are usually the highest-leverage. The antigravity-only finding
  on the stale dep would have been missed by single-lens review.

The disagreement table (rendered separately) shows which lens caught what. For security-audit, 1/N findings ARE the value — single-lens review misses cross-lens vulnerabilities.

Cache + collision prevention

Cache key is salted with the preset name, so a pr-review --pr 42 and security-audit --pr 42 produce distinct cache dirs:

• .kaijutsu/pr-review-runs/<sha>/ (pr-review of PR 42)
• .kaijutsu/security-audit-runs/<sha>/ (security-audit of PR 42)

Same diff can be reviewed by both presets without overwriting either's cache.

Composes

Chain naturally with:

• pr-review — ran first; security-audit is the security-specific deep dive when pr-review surfaced security-adjacent concerns
• decide — record any security tradeoffs surfaced (e.g., "we accepted medium-severity timing leak because mitigation cost > impact")
• dcg — already in the kaijutsu core; runtime guard for destructive shell commands. security-audit catches design-level issues; dcg catches execution-level ones.

Hard rules

• Never --post-comment for security-audit. The --post-comment flag is NOT registered on this subcommand — posting vuln details to a public PR is a leak. If you need to share findings, route them through your team's private channels.
• CVSS severities are not a substitute for triage. A "high" finding from security-audit is multi-model output, not an actual CVSS-scored vulnerability. Use the output as the seed for a real triage process, not as the final verdict.
• --allow-secrets is even more dangerous here than for pr-review. The audited code likely contains credentials by definition (auth flows, key handling). Consider redacting before running OR running on a sandboxed snapshot.

When NOT to use

• General code review — use pr-review.
• Spec/design review — use doc-review.
• One-line auth fix — overkill; manual review is fine.
• Compliance-grade audit — multi-model heuristic review is NOT a substitute for a real security audit by a human auditor. Use security-audit as a fast first-pass only.

Cost

Default --max-cost $1.00. With 3 agents in --full --strict (the default), typical security-audit runs ~$0.30–1.00 depending on input size. The debate round adds ~50% to base cost; the strict-filter adds another synthesis call. Override --max-cost higher for big audits.