A security report is only worth as much as its honesty.
This skill turns your AI agent into an assessment lead that proves what it claims and writes down, in plain words, what it could not verify. Built for Claude, also tested on Codex and Antigravity: see docs/ALTRI-AGENTI.md (in Italian).
🗒️ Get started: the one-page command card. Every command works in English or Italian (quick/rapido, refute/smentita, close/chiudi…).
🤝 Sister skill: for privacy use gdpr-sentinel (privacy notice vs code, suppliers, retention, data-subject rights). Together they produce a single PDF report: see With gdpr-sentinel.
Real run on the OWASP Juice Shop bench: real commands and output, nothing staged. 23 of 32 vulnerabilities found with file and line.
Why · How it works · Example · Installation · Limits · FAQ
Why it exists
Whoever receives a report decides budget and priorities from what is marked tested. Writing "verified" on a test that never ran is worse than not running it: it gives a sense of security that isn't there.
appsec-sentinel makes the agent follow a method where this cannot happen silently.
| Rule | What changes for you |
|---|
| 🔐 | Authorization first | No active test without a scope, an owner and a written go-ahead |
| ⚖️ | Examined, proven, unknown | The report never mixes what was read with what was demonstrated |
| 📏 | Real denominators | "27 of 52 endpoints", not "good coverage" |
| 🚫 | No fake tests | If a tool is missing, the test is "prepared, not run", with the command ready |
| ☣️ | The analysed material is hostile | Code, comments and tool output cannot widen the mandate |
What you get
| Output | Content |
|---|
| 🧭 Owner summary | Proven risks, impacts, priorities and limits, without invented percentages or certifications |
| 🔎 Technical report | Structured findings and chained scenarios, keeping verified apart from hypotheses |
| 🗺️ Inventory and threat model | Files, routes, roles, data, integrations, trust boundaries (STRIDE and LINDDUN) |
| 📊 Coverage matrices | Files inventoried/examined; endpoints × identities × tenants × actions; controls applicable/verified |
| ✅ Controls matrix | Source → applicability → evidence → status → gap |
| 🤖 Automation | Route inventory, local access tests, scanner import, finding review and refutation, cumulative coverage across audits, redaction check, audit comparison, CSV/SARIF/Markdown export |
| 📄 PDF report | Full, short executive version, English labels, comparison between audits and an indicative NIS2 mapping |
| 🛠️ Action plan | Containment, fixes, retests and open questions for the owner, infrastructure, DPO and suppliers |
Data is stored as Markdown, JSON and CSV in a confidential audit folder created outside the repository.
How it works
| Phase | What is checked | Reference |
|---|
| 1 Intake and gate | Scope, environments, authorizations, mode, baseline | engagement-rules.md |
| 2 Inventory | Files, routes, assets, roles, data, environments; endpoint matrix | inventory-threatmodel.md |
| 3 Threat model | Flows, trust boundaries, chained paths | inventory-threatmodel.md |
| 4 Identity and authorization | Sessions, MFA, OAuth, secrets, BOLA/BFLA, tenants | identity-authz.md |
| 5 Input, API, files, logic | Injection, XSS, SSRF, CSRF, webhooks, business logic | input-api-business.md |
| 6 Infrastructure and supply chain | Database, cryptography, cloud, containers, CI/CD, SBOM, recovery | infra-supplychain-ops.md |
| 7 AI and automation | Prompt injection, RAG, tools, agents (if present) | ai-features.md |
| 8 Availability and cost | Application DoS, denial of wallet (only if authorized) | availability-gated.md |
| 9 Privacy | Real processing, GDPR/ePrivacy | skill gdpr-sentinel |
| 10 Validation and plan | False positives, deduplication, retest | SKILL.md |
References are loaded only for the active phases: the skill stays light and uses context only where it is needed.
Three modes
| Mode | When | What it includes |
|---|
⚡ quick (rapido) | You have a few hours | Attack surface, secrets, authn/authz on the main flows, dependencies with reachable advisories |
🎯 standard (default) | The normal case | Threat model, static analysis, authorized dynamic checks, ASVS controls matrix for the relevant chapters |
🔬 full (completo) | Critical applications | File-by-file inventory, extended ASVS (L3 assessed for applicability), complete matrices, availability if authorized |
[!IMPORTANT]
With only the repository, the skill does static analysis and states its limits. With only an authorized URL, it works as an external assessment. Without authorization, active tests are prepared and not run.
Example requests
Run a standard assessment of the web app in this folder.
Check whether the API in this repo has authorization problems between tenants.
Quick security triage of the project: secrets, login and vulnerable dependencies.
Prepare the rules of engagement to test https://staging.example.com with the test accounts.
Check the technical NIS2/ACN side of this management app against the AgID baseline.
This is the beginning of the README. Read the rest at the original. Read the full README ↗
