AppSec Vulnerability Auditor
alboechat/appsec-vulnerability-auditor
Auditoria AppSec defensiva para código gerado por IA e vibe-coded
Ver no GitHubSkill defensivo de auditoria de segurança para Claude, especializado em análise de código gerado por IA e desenvolvido sem rigor formal. Produz relatórios priorizados de vulnerabilidades com mapeamento de superfície de ataque, tags CWE/OWASP, patches concretos e planos de remediação.
Instalação
npx convez-ai alboechat/appsec-vulnerability-auditor1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.
Métricas
Casos de uso
- Revisar código gerado por LLMs antes de deploy
- Auditar aplicações web quanto a vulnerabilidades OWASP Top 10
- Gerar relatórios de conformidade em SARIF, Markdown ou JIRA
- Identificar falhas de autorização e criptografia quebrada
- Detectar injeção de prompt em cadeias de ferramentas LLM
Veredito

Ferramenta robusta e bem-documentada para AppSec defensiva, com foco inteligente em falhas típicas de código gerado por IA. Recomendada para equipes que precisam auditar rapidamente código de origem duvidosa ou LLM-assisted, com saídas prontas para produção e planos de remediação estruturados.
README
A tradução para o português deste README ainda está na fila.
What this skill does
You give Codex or Claude code (a file, a diff, a PR, a directory). The skill produces a prioritized vulnerability report — Critical / High / Medium / Low / Informational — with:
- A short attack-surface map so you can sanity-check what was reviewed.
- Each finding tagged with CWE ID + OWASP category + file:line.
- Severity calibration for each finding: confidence, assumptions, downgrade/upgrade conditions, and validation questions.
- Language-specific checks for TypeScript/JavaScript, Python, Go, Rust, Java, Ruby, PHP, C#/.NET, and Swift/Kotlin mobile code.
- Product-ready outputs: Markdown, SARIF 2.1.0, GitHub issue bodies, JIRA tickets, and executive summaries.
- Repository-scale audit protocol with resumable
audit-statetracking. - Synthetic evaluation suite with 5 benchmark cases and 20 expected findings.
- Community and professional productization guidance for BioIntuitiva-backed use.
- A reproduction or sketch for every Critical/High (no weaponized PoCs).
- A concrete patch for every finding, in the same language and style as the original code.
- A list of what was checked and found clean (so the report is falsifiable).
- A "next steps" punch list ordered by severity.
- An optional Remediation Plan Builder (Step 10) that lets you choose which findings to fix, then generates a
.mdwith sprints, task checklists, code hints, acceptance criteria, and residual risk.
It is biased — on purpose — toward the failure modes that show up most in AI-assisted development: missing authorization checks, plausible-looking-but-broken crypto, prompt-injectable tool-use chains, hardcoded secrets, and server-trusts-client patterns.
What it looks for
The skill checks for standard web security risks, AI-generated-code failure modes, and LLM-agent-specific issues. Core categories:
Coverage summary:
- 16 major vulnerability classes documented below.
- 85 language/framework-specific checks across TypeScript/JavaScript, Python, Go, Rust, Java, Ruby, PHP, C#, and Swift/Kotlin.
- 100+ AppSec checks when combining core classes, language packs, OWASP Top 10, CWE Top 25, AI-generated-code pitfalls, and LLM-agent security risks.
These counts are audit-check coverage, not a promise of 80 distinct non-overlapping vulnerability types; several checks intentionally overlap across OWASP, CWE, and language-specific lenses.
| Category | Examples |
|---|---|
| Broken access control | Missing ownership checks, IDOR, tenant-boundary bypass, admin routes without policy checks |
| Authentication failures | Missing auth on protected routes, weak password storage, weak reset flows, session fixation |
| Authorization mistakes | Client-controlled roles, mass assignment of role / is_admin / tenant_id, privilege escalation |
| Injection | SQL/NoSQL/LDAP injection, OS command injection, template injection, unsafe dynamic queries |
| XSS and browser sinks | dangerouslySetInnerHTML, raw innerHTML, unsafe rich-text rendering, missing CSP |
| SSRF | User-controlled URL fetchers, PDF/rendering fetches, webhook fetchers, cloud metadata reachability |
| Unsafe deserialization | pickle, yaml.load, Java ObjectInputStream, PHP unserialize, Ruby Marshal.load |
| Cryptographic failures | MD5/SHA-1/SHA-256 passwords, ECB mode, hand-rolled JWT validation, weak randomness |
| Secrets and config | Hardcoded API keys, real values in .env.example, debug mode in production, insecure cookies |
| Server-trusts-client | Pricing, discounts, role, quantity, tenant, or ownership accepted from frontend state |
| File handling | Path traversal, unsafe upload/download, user filenames, missing content-type and size checks |
| Rate limits and abuse | No login/reset throttling, resource-exhaustion paths, unbounded request bodies |
| Error leakage | Stack traces, SQL errors, secrets or internals returned to users |
| Logging and monitoring gaps | No audit log on admin/security-sensitive actions, no off-host logs |
| Dependency / supply chain signals | Hallucinated packages, obvious vulnerable/deprecated dependency patterns, slopsquatting risk |
| LLM-agent security | Prompt injection into tools, indirect prompt injection, MCP tool authz, context exfiltration, output-as-input loops |
Language-specific packs add framework/runtime checks for:
- TypeScript / JavaScript (9 checks): Express/NestJS route protection, Next.js server actions, Prisma raw queries, JWT misuse, prototype pollution, frontend secret exposure.
- Python (9 checks): Django/Flask/FastAPI route auth, SQLAlchemy raw SQL, unsafe
pickle/yaml.load,subprocess(shell=True), debug config. - Go (9 checks): handler authz, SQL formatting,
os/exec,text/templatevshtml/template, unbounded body reads,InsecureSkipVerify. - Rust (10 checks):
unsafeboundaries, SQL string building, serde shape validation, path traversal, SSRF, resource exhaustion. - Java / JVM (9 checks): Spring
@PreAuthorize, mass assignment DTOs, JPQL/raw SQL, Jackson polymorphic typing, XXE, actuator exposure. - Ruby / Rails (9 checks): Pundit/CanCan gaps, strong-params bypass, ActiveRecord injection,
.html_safe,Marshal.load, Sidekiq argument risks. - PHP (10 checks): Laravel/Symfony policy gaps, raw SQL, file inclusion,
unserialize, CSRF, WordPress nonce/capability checks. - C# / .NET (10 checks): ASP.NET Core authorization, Entity Framework raw SQL, Blazor XSS, System.Text.Json deserialization, CORS, anti-forgery tokens, user secrets, certificate validation.
- Swift / Kotlin (10 checks): Keychain/Keystore insecure storage, certificate pinning, deep-link/universal link validation, WebView JS bridge, exported Android components, biometric auth bypass, hardcoded secrets.
What this skill does not do
- Write working exploits, weaponized PoCs, or shellcode.
- Produce reconnaissance or fingerprinting tooling.
- Bypass authentication or licensing on systems you don't own.
- Substitute for a professional pentest, a compliance audit, or a formal threat model.
If you ask for offensive output, the skill will decline and offer the defensive equivalent (a detection rule, a fix, a regression test).
Install
Option A — Personal skills folder
git clone https://github.com/alboechat/appsec-vulnerability-auditor.git
mv appsec-vulnerability-auditor ~/.codex/skills/
Option B — Claude / Cowork plugin / shared skills
Drop the folder into the directory your Cowork plugin reads from (typically ~/Library/Application Support/Claude/local-agent-mode-sessions/<...>/skills/ or your team's shared mount).
Option C — Sync to OneDrive / shared drive
Use the included script:
./tools/deploy-to-onedrive.sh
Edit the script's DEST variable if your OneDrive path differs from /Volumes/One Drive/Skills/appsec-vulnerability-auditor/.
Use
In any Codex or Claude conversation that has this skill available:
"Run the AppSec Vulnerability Auditor on this file." (paste/upload code)
"Audit this PR for vulnerabilities."
"Claude wrote this last night — can you security-review it before I deploy?"
The skill triggers automatically on natural-language phrasings like audit, security review, find vulnerabilities, is this safe to ship. See SKILL.md for the full trigger list.
What's in this repo
appsec-vulnerability-auditor/
├── SKILL.md # The skill prompt Codex/Claude reads
├── README.md # This file
├── README-pt-BR.md # Portuguese quickstart
├── LICENSE # MIT
├── AUTHORS.md
├── DISCLAIMER.md # Review limits and defensive-use boundaries
├── CHANGELOG.md
├── ROADMAP.md
├── CONTRIBUTING.md
├── CODE_OF_CONDUCT.md
├── SECURITY.md # How to report a vuln in THIS skill
├── .gitignore
├── agents/
│ └── openai.yaml # UI metadata for Codex/OpenAI surfaces
├── .github/
│ ├── CODEOWNERS
│ ├── ISSUE_TEMPLATE/
│ │ ├── bug_report.md
│ │ └── feature_request.md
│ ├── PULL_REQUEST_TEMPLATE.md
│ └── workflows/
│ └── ci.yml # Validation, link checks, evals, packaging
├── assets/
│ └── logo.svg
├── remediation_plan_builder_architecture.svg
├── docs/
│ ├── architecture.md # How the skill is organized
│ ├── runtime-architecture.md # Runtime, lifecycle, state, and eval pipeline
│ ├── vibe-coding-research.md # Research notes on AI-code failure modes
│ ├── productization.md # Community/pro positioning
│ ├── feature-spec-remediation-plan-builder.md # Design spec for Step 10
│ ├── threat-model-template.md
│ ├── glossario-appsec-pt-br.md
│ ├── release-checklist.md
│ └── github-release-template.md
├── references/ # What the skill consults during an audit
│ ├── owasp-top-10.md
│ ├── cwe-top-25.md
│ ├── ai-generated-code-pitfalls.md
│ ├── llm-agent-security.md
│ ├── severity-calibration.md
│ ├── output-formats.md
│ ├── finding-schema.json # Canonical Findings Ledger schema
│ ├── repository-audit-protocol.md
│ ├── evaluation-protocol.md
│ ├── remediation-plan-protocol.md # Step 10 plan builder protocol
│ ├── seed-input-protocol.md # SARIF/Semgrep/Gitleaks seed ingestion
│ ├── lang/
│ │ ├── typescript.md
│ │ ├── python.md
│ │ ├── go.md
│ │ ├── rust.md
│ │ ├── java.md
│ │ ├── ruby.md
│ │ ├── php.md
│ │ ├── csharp.md
│ │ └── swift-kotlin.md
│ └── secrets-and-config.md
├── templates/
│ └── remediation-plan.template.md # Structural template for plan output
├── examples/
│ ├── example-audit-report.md # Reference output the skill emulates
│ ├── findings-ledger.example.json
│ ├── remediation-plan-example.md # Example plan derived from notes-app audit
│ ├── anonymous-case-study.md
│ ├── repository-audit-summary.md
│ ├── audit-state.example.json
│ └── output-formats/
│ ├── sarif-example.json
│ ├── github-issue.md
│ ├── jira-ticket.md
│ └── executive-summary.md
├── evals/
│ ├── README.md
│ ├── eval-manifest.json
│ ├── scorecard-latest.json
│ ├── scorecard-template.md
│ └── cases/
│ ├── case-01-notes-api.md
│ ├── case-01-notes-api.json
│ ├── case-02-llm-agent-tools.md
│ ├── case-02-llm-agent-tools.json
│ ├── case-03-python-file-service.md
│ ├── case-03-python-file-service.json
│ ├── case-04-java-admin-api.md
│ ├── case-04-java-admin-api.json
│ ├── case-05-php-wordpress-plugin.md
│ └── case-05-php-wordpress-plugin.json
└── tools/
├── validate.sh # Lints repo (JSON, markdown frontmatter, bash)
├── check-skill-links.sh # Blocking SKILL.md link checker
├── run-evals.sh # Synthetic/live eval orchestrator
├── appsec-audit.py # Synthetic eval mock, not a real auditor
├── eval-compare.py # Eval comparator and scorecard writer
├── seed-with-semgrep.sh # Semgrep-to-ledger seed adapter
├── package-zip.sh # Builds release zip
├── deploy-to-onedrive.sh # Syncs to /Volumes/One Drive/Skills/...
└── github-init.sh # git init + gh repo create
Design philosophy
Three commitments, in tension on purpose:
- Calibrated severity over thoroughness theater. A report with five real Highs is more useful than one with twenty inflated Mediums. The skill is instructed to label honestly.
- Patches, not just findings. Every finding ships with a fix in the original style. A finding without a remediation is a complaint.
- Defensive only, no exceptions. The skill refuses to produce offensive tooling even when framed as "for testing." This is non-negotiable — see
SECURITY.md.
Evaluation
The evals/ suite contains five defensive cases with 20 expected findings. The shipped harness runs in synthetic mode by default; it validates the comparator and scorecard pipeline, not a real LLM's security skill. Use a future/live appsec-audit runner to measure recall, precision, severity accuracy, calibration completeness, and safety against a model. See references/evaluation-protocol.md and evals/README.md.
Community and professional use
The public repository is the community edition: self-serve skill, references, outputs, repository protocol, and evals. BioIntuitiva can build professional services around it: scoped reviews for AI-built MVPs, threat-model intake, validated findings, fix plans, re-tests, and executive summaries. See docs/productization.md.
Why "vibe-coded"?
"Vibe coding" is the practice of building software by describing intent to an LLM and accepting most of what it writes. It is enormously productive and produces a recognizable, repeatable set of security failures — see docs/vibe-coding-research.md. This skill is tuned for those failures because they are the ones professional auditors are now seeing the most of, in the highest-volume codebases, with the lowest baseline review.
Contributing
Issues, PRs, and reference-list expansions welcome. Please read CONTRIBUTING.md and CODE_OF_CONDUCT.md first. To report a vulnerability in this skill itself, follow SECURITY.md.
Author
Arquitero/Engenheiro Prof Dr Antonio Luiz Boechat, BioIntuitiva.
License
MIT — see LICENSE.