security-review

name	security-review
description	Manual-only skill. Use ONLY when the user explicitly invokes: $security-review Never select this skill via semantic matching.

Security Review

Activation guard

If "$security-review" is NOT present in the user request:

• Exit immediately
• Do not analyze anything
• Do not load references

---

Token policy

• Default: no reference files
• Load at most ONE reference file
• Only load references if strictly required
• Prefer repo evidence over general guidance
• Keep output concise and high-signal

---

Modes

Light mode (default)

Use for:

• narrow code paths
• quick checks
• implementation guidance

Rules:

• no references unless absolutely necessary
• only high-confidence findings
• minimal output

---

Full review mode

Use ONLY if explicitly requested (e.g. “full security review”, “threat model”)

Rules:

• define scope first
• inspect repo before references
• load at most one reference initially

---

Workflow

1. Identify scope (files, components, boundaries)
2. Locate:
   • entrypoints
   • auth boundaries
   • user input paths
   • sensitive data handling
3. Analyze only what is in scope
4. Escalate only if risk justifies it

---

Priority risks

Focus on real, high-impact issues:

• auth / authorization flaws
• injection risks
• SSRF / request forgery
• unsafe parsing / deserialization
• file upload / path traversal
• secrets exposure
• tenant isolation failures
• insecure defaults

Ignore:

• style issues
• speculative edge cases
• generic checklists

---

Output format

For each issue:

• Issue: what is wrong
• Location: where it is
• Impact: why it matters
• Abuse path: how it could be exploited
• Fix: smallest safe remediation

If no issues: say so clearly.

---

Reference usage

Only if necessary, choose ONE:

• golang backend
• browser JavaScript / HTMX frontend
• security controls and assets

If no strong match → use none

---

Hard rules

• Only run when "$security-review" is present
• Never trigger implicitly
• Never scan entire repo unless explicitly asked
• Never load multiple references unless absolutely required
• Never inflate output with generic advice