| name | cybersecurity-audit |
| description | Stack-agnostic vulnerability hunting — OWASP Top 10 2025, OWASP API Top 10, OWASP LLM Top 10, supply-chain via OSV.dev + native scanners, secrets, crypto, web (XSS/CSRF/CORS), SSRF, deserialization/RCE, broken access control, race conditions. Produces categorized findings + remediation plan. Use when: проверь на уязвимости, security audit, найди дыры, аудит безопасности, OWASP, SQL injection, XSS, CSRF, IDOR, SSRF, prompt injection, supply chain, dependency CVE, hardcoded secrets, security review, penetration testing prep. SKIP: secure-coding discipline upfront (→karpathy-guidelines); SOC2/PCI-DSS/HIPAA compliance audits; live exploitation / payload generation (defensive-only). |
| tags | ["security","vulnerabilities","owasp","audit","supply-chain","osv","llm-security"] |
Usage
Loaded automatically when the user asks for a security audit, vulnerability check, or OWASP review. Used by the worker-security-verifier agent for its ## Vulnerability sweep step. Stack-agnostic — works on Node/Python/Astro/Next/CRM/Telegram-bot/anything.
Purpose
Most security scanners check one ecosystem and one class of issue. Real attackers chain: leaked secret → SSRF → IMDS credentials → S3 read → user data. A real audit covers all classes, in all stacks, with a remediation plan that prioritizes by exploitability × blast-radius, not just by CVSS score. This skill is the curated 2026 knowledge base for that.
Use this skill when
- User asks to find vulnerabilities in code/codebase
- After dependency bump → verify CVE picture
- Before launching to public → final security pass
- After incident → root-cause-and-radius assessment
- When refactoring auth, crypto, payments, file uploads, external API calls
- When integrating a new third-party SDK (supply-chain check)
- When LLM features are added (prompt injection surface)
Do not use this skill when
- Writing new code from scratch — use
karpathy-guidelines to write safely upfront
- Compliance reports (SOC 2 / ISO 27001 / PCI-DSS / HIPAA) — niche; out of scope
- Active exploitation / penetration testing with real payloads — this skill is defensive-only: find issues + recommend fixes, never craft live exploits
Capabilities
OWASP Top 10 2025 — the foundation
The 10 most common web vulnerability classes, refreshed every few years by OWASP from real-world breach data. Apply this taxonomy to ANY web/API/service code:
- Broken Access Control
- Cryptographic Failures
- Injection (SQL/NoSQL/Command/LDAP/XPath/SSTI)
- Insecure Design
- Security Misconfiguration
- Vulnerable & Outdated Components
- Identification & Authentication Failures
- Software & Data Integrity Failures
- Security Logging & Monitoring Failures
- SSRF
→ Deep dive: references/owasp-top-10-2025.md
OWASP API Top 10 — for REST/GraphQL surfaces
Separate taxonomy because API attacks differ from form-based webapps: BOLA (Broken Object-Level Authorization), mass assignment, GraphQL introspection in prod, rate-limit absence, etc.
→ Deep dive: references/owasp-api-top-10.md
OWASP LLM Top 10 — for AI features
Specific to LLM apps: prompt injection (direct/indirect), training-data poisoning, model DoS, sensitive info disclosure, system-prompt leak, insecure output handling.
→ Deep dive: references/owasp-llm-top-10.md
Injection patterns (deep)
Every injection vector with detection grep + remediation pattern: SQL (parameterize), NoSQL ($where), OS command, LDAP, XPath, Server-Side Template Injection (SSTI), header injection.
→ Deep dive: references/injection-patterns.md
Auth, session, access control
JWT pitfalls (alg: none, weak secret, missing exp), session fixation, IDOR, missing authz checks, privilege escalation, OAuth misconfig.
→ Deep dive: references/auth-access-control.md
Crypto + secrets
Weak hashes (MD5/SHA1 for passwords), broken modes (ECB), IV reuse, predictable RNG (Math.random() for tokens), timing-side-channels, hardcoded keys/tokens, .env hygiene, gitleaks patterns.
→ Deep dive: references/crypto-secrets.md
Supply chain — OSV.dev + native scanners + typosquatting
This is where OSV.dev lives. Cross-ecosystem CVE check via Google's Open Source Vulnerabilities database, plus native scanners (npm audit, pip-audit, cargo audit, govulncheck), plus checks for typosquatting and dependency confusion.
→ Deep dive: references/supply-chain-osv.md
→ Script: scripts/run-osv-scan.sh
Web frontend — XSS / CSRF / CORS / CSP
DOM/Stored/Reflected XSS, CSRF tokens vs SameSite, CORS misconfig (Access-Control-Allow-Origin: * with credentials), CSP holes (unsafe-inline, unsafe-eval), clickjacking, open redirect.
→ Deep dive: references/web-frontend.md
SSRF, deserialization, RCE
Untrusted-input-to-code-execution chains: SSRF → IMDS / internal network; pickle/eval/Function/exec on user input; YAML.load unsafe; XML XXE.
→ Deep dive: references/ssrf-deserialization.md
Race conditions
TOCTOU (Time-of-Check-Time-of-Use), idempotency window in payments (double-spend), token reuse, concurrent state mutation.
→ Deep dive: references/race-conditions.md
Remediation plan
Findings → priority queue. Severity (Critical/High/Medium/Low) × exploitability × blast-radius → ordered checklist with owners + ETAs.
→ Template: references/remediation-plan.md
Behavioral Traits
- Defaults to NEEDS WORK. A green sweep means every check in the matrix ran clean — not "I ran a few and nothing jumped out".
- Assumes hostile input at every boundary (HTTP body, query params, headers, file uploads, external API responses, env vars from untrusted sources).
- Treats secrets as already leaked — checks rotation cadence and blast radius, not just presence.
- Reads diffs first, then full files — incremental audit is cheap, full-file rescan is expensive.
- Outputs remediation — not just "found XSS"; says "found XSS in
routes/user.ts:42; fix: replace innerHTML = req.body.name with textContent or escape via DOMPurify".
- Prioritizes by exploitability + blast — a Critical CVE in a dev-only dep can wait; a Medium IDOR in user data cannot.
- Cross-ecosystem checks via OSV.dev — does not trust a single native scanner.
- Reports INCONCLUSIVE when a check can't be run, never fakes PASSED.
- Never auto-fixes — defensive-only; surface findings, suggest, let humans (or main thread) apply.
Important Constraints
- NEVER craft live exploit payloads (curl injections, RCE PoCs against real targets) — this is defensive-only
- NEVER run
npm audit fix / pip-audit --fix / any auto-remediation
- NEVER mark a finding as "false positive" without code evidence (file:line + reasoning)
- NEVER skip a check because "this stack probably doesn't have X" — run the check, let the result speak
- NEVER commit findings into the codebase (don't leave annotated source as a record — write into a separate report)
- ALWAYS cross-check supply-chain via OSV.dev when native scanners report 0 findings (sometimes they miss)
- ALWAYS rate severity using CVSS-aligned criteria + business context, not personal feel
- ALWAYS surface secrets findings with full file path + git blame so user knows when/who introduced
Related Skills
Authoring layer
- ✓
karpathy-guidelines — write secure code from the start (prevention)
- ✓
skill-evaluation — for editing this skill
Stack skills (this skill references their conventions)
- ✓
better-auth — auth check patterns
- ✓
zod — schema validation strictness
- ✓
pydantic — same for Python
- ✓
postgresql — SQL injection prevention via parameterization
- ✓
redis — sensitive key TTL + ACL
- ✓
nodejs — process.env hygiene + AbortController
- ✓
cloudpayments · yookassa — payment-specific HMAC + idempotency
- ✓
telegram-bot · vk-bridge · max-bridge — initData/launch-params HMAC validation
Tools the skill recommends running
osv-scanner (Google) — npx osv-scanner -L package-lock.json or osv-scanner --recursive .
semgrep ci --config=auto — pattern-based static analysis (free)
gitleaks detect --redact — secret scan
trivy fs . — IaC + container + deps
bandit -r . — Python static
npm audit / pip-audit / cargo audit / govulncheck — ecosystem-native
API Reference
| Topic | File |
|---|
| Index, decision map, what-to-open-when | references/REFERENCE.md |
| OWASP Top 10 2025 — full breakdown with grep patterns | references/owasp-top-10-2025.md |
| OWASP API Top 10 — REST/GraphQL-specific | references/owasp-api-top-10.md |
| OWASP LLM Top 10 — prompt injection, system-prompt leak, etc. | references/owasp-llm-top-10.md |
| Injection patterns — SQL/NoSQL/Command/LDAP/XPath/SSTI | references/injection-patterns.md |
| Auth, sessions, JWT, IDOR, missing authz | references/auth-access-control.md |
| Crypto pitfalls + secrets leakage + .env hygiene | references/crypto-secrets.md |
| Supply chain — OSV.dev + osv-scanner + native scanners + typosquatting | references/supply-chain-osv.md |
| Web frontend — XSS / CSRF / CORS / CSP / clickjacking / open redirect | references/web-frontend.md |
| SSRF + deserialization + XXE + RCE chains | references/ssrf-deserialization.md |
| Race conditions — TOCTOU, idempotency, double-spend | references/race-conditions.md |
| Remediation plan template — Critical→Low ordered checklist | references/remediation-plan.md |
Templates
Scripts