Security Check

Review the current change for security issues. You are a thorough but pragmatic reviewer — focus on real risks, not theoretical edge cases. Do not edit files.

Workflow

1. Run git diff main...HEAD (or git diff for uncommitted) to see what changed.
2. Launch the cross-agent pass in the background if it qualifies (see below) — before forming any findings of your own.
3. Read modified files in full where context matters.
4. Check against OWASP Top 10 + STRIDE.

Cross-agent pass (opportunistic)

Security is exactly the category where a second model pays for itself — and the diff is often authored by the model now checking it.

Which CLI am I running in? Use only the matching branch below; never call your own CLI.

• In Claude Code: your model is Claude and the other model is Codex. It is available if codex --version succeeds. Launch it with:

  codex exec --cd "$PWD" --sandbox read-only - < "$PROMPT_FILE" > "$OUT_FILE" 2>&1 &

• In Codex: your model is Codex and the other model is Claude. It is available if claude --version succeeds. Launch it with:

  claude -p --permission-mode plan < "$PROMPT_FILE" > "$OUT_FILE" 2>&1 &

Then, in either CLI, if the other model is available and the diff is non-trivial (≥ 20 changed lines):

1. Right after computing the diff (step 1), write a prompt file containing:

   • the diff range
   • the full OWASP + STRIDE checklist below
   • the findings format — without the provenance tags or the dual/single-model label line; you add those at merge
   • a read-only instruction: do not edit files or run write operations

   Not your own findings — the other model must check blind, or it will anchor on your framing. Launch it in the background with your branch's command, capturing output to a file (docs/cross-agent-review.md has the canonical background form; prefer your harness's background facility over bare & if it has one).

2. Do your own full pass and draft your own findings before reading the other model's output.
3. After your draft is done, wait up to ~90 seconds for the other model; if it hasn't returned, proceed single-model and say so.
4. Merge with provenance: tag every finding [both], [claude], or [codex]. Verify findings only the other model raised ([codex] in Claude Code, [claude] in Codex) against the actual code before including them. Findings tagged [both] carry the most confidence, but confidence is not priority: a single verified vulnerability outranks two matching low-severity notes. Add one DISAGREEMENTS: line per explicit contradiction; silence from a model is not disagreement.

This pass never blocks the check: if the other model is missing or slow, complete single-model and say so in the SECURITY SUMMARY.

OWASP Top 10 pass

For each, note: present / not applicable / concern.

1. Broken access control — anyone able to access something they shouldn't?
2. Cryptographic failures — anything sensitive transmitted or stored in the clear; weak algorithms; hardcoded keys?
3. Injection — SQL, command, LDAP, OS, prompt injection?
4. Insecure design — missing rate limits, no auth where needed, trust boundaries violated?
5. Security misconfiguration — overly permissive CORS, verbose error messages, default creds?
6. Vulnerable & outdated components — new deps; any with known CVEs?
7. Identification & authentication failures — weak session handling, missing MFA, token leakage?
8. Software & data integrity failures — unsigned updates, deserialization issues, dependency tampering risk?
9. Security logging & monitoring failures — enough audit trail for what was added?
10. Server-side request forgery — user-controlled URLs being fetched server-side?

STRIDE pass

• Spoofing — can someone impersonate a user or service?
• Tampering — can data in transit or at rest be modified undetected?
• Repudiation — can a user deny doing something we'd need to prove?
• Information disclosure — anything leaked through errors, logs, response bodies?
• Denial of service — unbounded loops, unrate-limited endpoints, easy resource exhaustion?
• Elevation of privilege — can a low-privilege actor get higher access?

Output

SECURITY SUMMARY: [one line — overall risk read]
[dual-model check (<your model> + <other model>) | single-model check (reason)]

HIGH RISK (block merge):
- [both|claude|codex] [file:line] [issue] → [fix]

MEDIUM RISK (fix before prod):
- [both|claude|codex] [file:line] [issue] → [fix]

LOW / NIT (track for later):
- [both|claude|codex] [file:line] [issue] → [fix]

CHECKED & CLEAN: [list of OWASP/STRIDE categories that were checked and are clean — to show coverage]

Omit the provenance tags when the check was single-model.

Rules

• Don't cry wolf. A theoretical attack with no realistic path doesn't belong in HIGH RISK.
• For each finding, the fix matters as much as the identification. "This is broken" without a remediation is half a finding.
• If you didn't check something (e.g., infra not in scope), say so — don't claim coverage you don't have.