security

name	security
description	Security review skill for threat modeling, dependency audit, secret hygiene, and code review through a security lens. Activate when a change touches authentication, authorization, secrets, cryptography, input parsing, deserialization, network calls, or upgrades to dependencies.
type	skill
license	Apache-2.0
metadata	{"provenance":{"canonical":"${CANONICAL_REPO}","feedback":"${FEEDBACK_REPO}","version":"1.1.0"}}
claude	{"allowed-tools":["Read","Grep","Glob","Bash"],"user-invocable":true}

Security

A focused review skill, not an exhaustive checklist. The goal is to catch the realistic next exploit — not to enumerate the OWASP Top 10 every time.

When to activate

Mandatory triggers:

• Authentication, authorization, or session-handling code changes.
• Secret / credential / token storage or transmission paths.
• Cryptographic primitive use (hashing, signing, encryption, RNG).
• External input parsing (HTTP bodies, CLI args, file uploads, IPC).
• Deserialization of untrusted data.
• Outbound network calls (URLs from user input, redirects, SSRF surface).
• Dependency additions or upgrades touching the surface above.

Optional but encouraged: any code review on a security-adjacent module (payment, PII, audit log) — even when the diff itself looks innocuous.

Operating mode

1. Trust boundary first

Before reading the diff, draw the trust boundary in your head:

• What input is untrusted here? (User input, network, file from another process, env var passed in by the user.)
• What is inside the boundary? (Internal config, signed messages, vetted constants.)
• The bug class to look for is the one that crosses the boundary without proper handling.

2. Realistic threat, not paranoid theatre

Ask: what would the next attacker actually try here? Two concrete threats with credible exploit paths beat a list of twenty generic risks. Examples of credible threats per surface:

• HTTP handler reading from JSON: prototype pollution, type confusion, unbounded length / depth.
• File path from user: traversal, symlink, race.
• SQL: parameterised? If yes, also check ORDER BY / LIMIT / table name interpolation, which most ORMs do not parameterise.
• Crypto: AEAD with reused nonce, hash truncation, padding oracle, RNG not from a CSPRNG.
• Deserialization: gadget chains, auto-instantiation.

3. Verify, do not assume

For each suspected issue, verify by:

• Reading the relevant function end-to-end (not just the diff hunk).
• Tracing the data flow from boundary to sink.
• Checking the test suite — if no test exists for the threat path, flag the absence as part of the finding.

Do not flag a "potential" issue without showing how the data could flow there. Speculation costs the user time and erodes credibility.

4. Output format

Report findings as a numbered list:

1. [SEV] <one-line summary>
   Where: <file:line>
   Threat: <what an attacker would do>
   Evidence: <code snippet or trace>
   Fix: <concrete recommendation>

Severity scale: BLOCKER / HIGH / MED / LOW / INFO. Reserve BLOCKER for issues that must be fixed before merge (e.g. credential leak, RCE, auth bypass).

Secret hygiene

• Never commit .env, credentials.json, *.pem, or any file under a vault path.
• Never echo a secret in logs, even at debug level. Even if the log is ephemeral, the transcript may be persisted.
• If you find a leaked secret in the repo or transcript, rotate guidance is part of the finding — flag the secret as compromised the moment it touches a non-secret-grade store.

Dependency audit

When asked to review a dependency upgrade:

• Read the upstream changelog between current and target versions.
• Check the GitHub Security Advisory database for known CVEs.
• For minor / patch bumps, focus on transitive changes and licence drift.
• For major bumps, flag breaking changes and require a migration note.

Friction reporting

When a recognition signal fires (see config/TOOLS.md → Friction Reporting → Recognition signals), invoke the harness-report skill rather than reimplementing the protocol inline. The reporter walks you through identifying the offender, picking the room, and filling the payload.