Security review

Security Review

Use this portable skill for non-Claude agents when the user asks for a security review, vulnerability audit, hardening pass, or security-focused PR/code review. On Claude Code, prefer the bundled /security-review command when it is available.

This skill is adapted from Anthropic's MIT-licensed anthropics/claude-code-security-review template for portable Overdrive runtimes.

Scope First

1. Determine whether the user wants a diff review, whole-repo audit, file-specific review, or threat-model pass.
2. For diff reviews, compare against the base branch if available.
3. For whole-repo reviews, prioritize entry points, auth boundaries, data flows, server actions/API routes, database access, file operations, secrets handling, and deployment config.
4. Do not perform destructive actions, live exploitation, credential use, account changes, or production probing without explicit approval.

Review Categories

Look for concrete, exploitable issues in:

• Injection: SQL, NoSQL, shell command, template, XML/XXE, path traversal, unsafe deserialization, YAML/pickle, eval/dynamic code.
• Web security: reflected/stored/DOM XSS, CSRF where relevant, unsafe redirects with real impact, SSRF where host/protocol is attacker-controlled.
• Authentication and authorization: bypasses, privilege escalation, missing server-side checks, session/JWT flaws, tenant isolation breaks.
• Data exposure: sensitive logging, debug information exposure, PII leaks, overly broad API responses, insecure public storage.
• Crypto and secrets: weak crypto, bad randomness, disabled certificate validation, hardcoded credentials, leaked tokens.
• Configuration and deployment: unsafe CORS, public admin/debug surfaces, insecure defaults, missing environment separation.
• Supply chain: suspicious install scripts, unpinned high-risk dependencies, dependency execution paths, untrusted package sources.
• Business logic: purchase, billing, quota, invite, ownership, or workflow bypasses with real user/security impact.

False-Positive Filter

Report only findings with a credible exploit path and meaningful impact. Skip:

• Style issues, refactors, or general code quality.
• Pure denial-of-service, rate-limit, or resource exhaustion concerns unless the user explicitly asks.
• Theoretical "could be safer" findings without a concrete attacker path.
• Documentation-only issues.
• Test-only code unless tests run privileged production-affecting actions.
• Client-side missing auth checks when the server correctly enforces authorization.
• React/Angular XSS concerns unless unsafe HTML APIs are used.
• Outdated dependency notices unless the task is specifically a dependency/security update.
• Log spoofing or non-sensitive logging.

Output

Lead with findings, ordered by severity. If there are no concrete issues, say so plainly and name the residual risk or test gaps.

Use this format:

## Findings

### High: <short title>
- Location: `path/to/file.ext:line`
- Category: <injection/authz/xss/data-exposure/etc>
- Why it matters: <impact>
- Exploit path: <specific scenario>
- Recommendation: <specific fix>
- Confidence: <0.8-1.0>

## No Findings
No high-confidence security issues found in the reviewed scope. Remaining risk: <limits of review>.

Verification

When practical and safe, verify fixes with tests, targeted static checks, dependency scanners, or minimal local reproduction. Do not invent proof you did not run.

Attribution

Adapted from Anthropic's MIT-licensed anthropics/claude-code-security-review repository: https://github.com/anthropics/claude-code-security-review