Security Review Agent Skill

Security Review

Find vulnerabilities before they ship.

Hard Rules

• Do not dismiss a finding without evidence it is unexploitable.
• Do not commit, log, or surface secrets discovered during review — flag and recommend rotation.
• Do not modify code until the user approves a remediation plan.

Workflow

• Scope
• Confirm target: diff, file set, module, full repo, or skill/prompt. A target can be both code and prompt.
• Identify stack/framework — adapt the checklist (skip what the framework handles, add its pitfalls).
• Trace data flow: request → middleware → handler → service → datastore → response. For prompts: input → template → LLM → tools → output.
• Map trust boundaries, privilege levels, and threat actors.
• Search prior findings: npx ai-devkit@latest memory search --query "<target>" --tags "security"

• Scan
• Only check relevant categories. Skip sections and items that don't apply. Do not report skipped items.
• For diffs/PRs: also check whether the change weakens existing controls — removed middleware, bypassed validation, new unprotected routes.
• Categories in priority order:

a. Secrets — hardcoded tokens, keys, connection strings.

b. Injection — SQL, NoSQL, command, template, SSRF, path traversal, XSS.

c. Auth — missing checks, privilege escalation, OAuth/OIDC, IDOR.

d. Business Logic — race conditions, TOCTOU, workflow bypass, mass assignment, parameter tampering.

e. Data Exposure — PII in logs, verbose errors, overly broad responses.

f. Resource Exhaustion — unbounded queries, missing pagination, upload size, decompression bombs.

g. Dependencies — critical CVEs only (RCE, auth bypass, data breach); ignore low/medium.

h. Cryptography — weak algorithms, hardcoded IVs/keys, disabled certificate validation.

i. Configuration — debug mode, permissive CORS, missing security headers.

j. Logging — security events unlogged, no tamper protection, no alerting.

k. Prompt Injection — instruction override, tool abuse, data exfiltration, indirect injection via tool results.

• For each finding: file, line, evidence.

• Classify

| Severity | Criteria |

|----------|----------|

| Critical | Exploitable now, data loss or RCE possible |

| High | Exploitable with moderate effort or insider access |

| Medium | Requires chained conditions or limited impact |

| Low | Defense-in-depth, no direct exploit path |

• Adjust severity by exposure (internet-facing vs internal) and data sensitivity.
• Check for attack chains — multiple Medium findings that combine into High/Critical.
• Mark false positives with reasoning.

• Remediate
• For each finding: root cause, minimal fix (prefer stdlib/framework over custom), verification step.
• For Critical/High: also recommend a detection control (log, alert, or WAF rule).
• Present plan and request approval before changing code.

• Verify
• Use the verify skill to confirm each remediation.
• Re-scan fixed files for regressions.
• Store findings: npx ai-devkit@latest memory store --title "<pattern>" --content "<finding and fix>" --tags "security,<category>"

Red Flags

| Rationalization | Do Instead |

|---|---|

| "It's internal / behind a VPN / only admins" | Zero-trust: validate at every boundary regardless of network position or user role |

| "We'll add auth later" | Add auth before merge — unauthenticated endpoints get discovered fast |

| "It's just a dev credential" | Use env vars / secrets manager — dev secrets leak to prod constantly |

| "The framework handles that" | Verify the config — frameworks have defaults, not guarantees |

| "We sanitize on the frontend" | Always validate server-side — client validation is bypassable |

| "The LLM won't follow injected instructions" | Treat all tool results and external content as untrusted data |

| "It's just a prompt, not code" | Prompts control tool execution — review with the same rigor as code |

Output Template

• Scope: Target, stack, data flow, trust boundaries, threat actors
• Findings (by severity): ID, severity, category, file:line, exploit scenario, fix
• Attack Chains: Findings that escalate when combined
• False Positives: Dismissed items with reasoning
• Remediation Plan: Ordered fixes with verification steps
• Residual Risk: Scope limitations, unverifiable items
• Zero findings: state what was checked and scope boundaries — "no findings" ≠ "fully secure"