mbart75/appsec-skills

Security skills for AI coding agents: design security in before coding, review code the way a security engineer would, fix findings one proven step at a time, and check French legal requirements before launch. Every output is a document that any other AI, or a person, can pick up.
Works with any agent that reads the open SKILL.md format: Claude Code, Codex, Cursor, GitHub Copilot, Gemini CLI, OpenCode. Chat assistants get a single-file version.
## What you get
Four skills that cover the life of a project, from design to launch:
| Skill | When | What it produces |
| appsec-threat-model | Before building an app or a feature (security by design) | docs/security/threat-model.md and requirements.md in the project: trust boundaries, concrete threats, testable requirements, each with how to verify it |
| appsec-review | While coding (diff), at any time (full), before going live (pre-launch), after fixes (verify); also on a deployed app without its source, with authorized tests | A verified report written as a work order: stable IDs, evidence with secrets redacted, attack scenario, fix, "Done when" check, status |
| appsec-harden | To fix a report's findings or apply a baseline (CI, headers, cookies, repository hygiene) | One commit per fix, each proven by a test, scanner or static check; rotations and decisions left explicitly to a person |
| appsec-compliance-fr | Before launching a site or app in France | A checklist with sources (mentions légales, RGPD, cookies, consumer law, reviews, AI Act, accessibility) and draft pages in French with placeholders |
The loop: appsec-threat-model writes the requirements, appsec-review checks the code against them, appsec-harden fixes what the review found, appsec-review verify confirms. Any assistant can take over at any step: the documents and the report carry everything it needs.
## Principles
The skills send no source code anywhere unless it is already public: no scanner that uploads code is used, and the scanner runner never uploads it. What the agent itself sends to its model provider is up to the agent you use; a skill cannot change it.
Scanners are optional. The skill runs the ones you have (gitleaks, TruffleHog, Semgrep, Bandit, ShellCheck, PSScriptAnalyzer, OSV-Scanner, Checkov, Trivy, zizmor), lists the missing ones, and never installs anything. Without any, it still works: most serious flaws are found by reading the code.
Known vulnerabilities are ranked by real-world exploitation, not by CVSS alone: each CVE found in dependencies, including vendored libraries, is checked against the CISA KEV catalog (exploited in the wild) and its EPSS score (likely to be exploited), then against the code (is the vulnerable function used?). To do this, OSV-Scanner sends package names and versions to the OSV database, and the ranking sends CVE identifiers to the EPSS API and downloads the KEV catalog. Semgrep and Trivy download their rules. --offline stops all of it: OSV-Scanner then uses a local database, downloaded beforehand with osv-scanner scan source --offline-vulnerabilities --download-offline-databases, and without one the summary says NO DATABASE, since nothing was checked; Semgrep, whose rules come from its registry, is skipped (SKIPPED (offline)); Trivy uses the checks embedded in its binary. Trivy is always run with --disable-telemetry and --skip-version-check, online or offline; a test runs the real Trivy behind a recording proxy and checks that it opens no connection with --offline. Online, the test checks the options passed, not Trivy's traffic.