security
- Group: Security
- When it activates: Reviewing or auditing the project’s security, hardening auth / an API / an AI agent, hardening, authorized pentesting, OWASP, prompt injection, XSS/CSRF/SSRF/IDOR, security headers
- Requires:
code-quality - Installed: Always: comes with every project.
Make the application withstand whoever attacks it: review it, find the holes with PROOF and close them. It does not
replace code-quality (which already carries the per-vulnerability detail); this skill is the security GATEWAY:
the review method, the pentesting policy and the hardening of AI agents. The command is /security.
Minimum reading per task
Section titled “Minimum reading per task”| Task | Read only |
|---|---|
Review/audit the project’s security (the WHAT that /security runs) |
references/metodo-revision.md |
| Pentesting, scanning or exploitation (scope, authorization, the wall) | references/pentesting.md |
| Harden an AI agent / LLM app (prompt injection, tools) | references/agentes-ia.md |
| Per-vulnerability detail (injection, XSS, CSRF, SSRF, uploads…) | code-quality §references/security-owasp.md |
| Login, sessions, JWT, OAuth, 2FA, password reset | code-quality §references/auth-patterns.md |
| Architecture of an LLM app (RAG, streaming, costs, evals) | code-quality §references/llm-apps.md |
| Secrets, environments, leaks and rotation in deployment | deploy-ops §references/envs-secrets.md |
| Supply chain (packages/ skills with malware or typosquatting) | /dependencies command + supply-guard wall |
Hard principles (always apply)
Section titled “Hard principles (always apply)”- Authorize on the server, on every operation, on the specific resource. Access control is the number 1 failure (OWASP A01). A menu hidden in the UI is not authorization.
- Validate all input on the server with an allowlist (schema), fail closed, least privilege and minimal surface. Client-side validation is UX only.
- Secrets never in the repo, in logs, in the client bundle or in a prompt. Leaked secret = rotate it NOW
(then clean the history).
.env.examplewithout real values. - Every finding with EVIDENCE: where it is (file:line or request), why it is exploitable and how it is fixed. Without proof it is not a finding, it is a hunch. When it is closed, it is re-verified.
- User input and what an LLM retrieves are DATA, not instructions. Real privileges live
outside the model: tools with an allowlist and the real user’s authorization; human confirmation for anything
irreversible (see
references/agentes-ia.md). - Pentesting ONLY with written authorization and a defined scope. The
pentest-guardwall stops offensive tools (nmap, sqlmap, hydra, metasploit…); theSENZU_ALLOW_PENTEST=1escape is for an authorized engagement documented in the devlog (references/pentesting.md). - Defensive security by default. You build to resist attacks and test your own systems with permission; no malware, no exploits for third parties, no detection evasion with malicious intent.
Relationship with other skills and walls
Section titled “Relationship with other skills and walls”code-quality carries the per-vulnerability detail (OWASP, auth, LLM) · backend-audit audits architecture and
debt · the /dependencies command + the supply-guard wall cover the supply chain · deploy-ops deployment
security (secrets, access) · walls that already protect: secrets-guard (secrets in code),
backend-guard (mass-assignment, PII in logs), pentest-guard (pentesting scope).