Claude Security plugin
Runs Anthropic's agent panel over a repository or a diff: researchers propose vulnerabilities, three verifiers try to disprove each one, and the report ships with SARIF and optional verified patch files. Use for a deep scan of your own code.
claude-security- Author
- Anthropic
- Source
- anthropics/claude-plugins-official/plugins/claude-security/skills/claude-security @ d7d3279
- License
- Proprietary, all rights reserved
About this skill
What it does
Claude Security is a front desk for three jobs on your own repository. A codebase scan runs as the plugin's scan workflow: an inventory splits the tree into components, a threat model goes first, then one researcher per component and attack category and a breadth sweep, plus a secrets pass on a large tree. Effort is low (one researcher plus the panel), medium (up to 24 components) or high (up to 48, two researchers per cell). A scan of changes points the same machinery at a branch diff, a pull request's diff or one commit, reporting only flaws the change takes part in. Every candidate goes to a three-voter panel whose job is to disprove it; survivors land in a timestamped CLAUDE-SECURITY-<ts>/ directory as a Markdown report, JSONL, a SARIF 2.1.0 log and a revision stamp, behind that directory's own .gitignore. A finding carries impact, exploit scenario, preconditions, severity and confidence, and the verification tally is computed in code, not asserted by the model.
"Suggest patches" develops each fix in a scratch clone, so your checkout is never touched. A patch file appears only when an independent verifier vouches that the change is targeted, adds no new vulnerability and leaves behaviour otherwise unchanged; otherwise you get a note saying why. Nothing is applied, committed or pushed. Besides the skill the plugin ships eight subagents, the scan.js workflow, Python helpers for the report, SARIF, CWE mapping and the revision stamp, and hooks that offer a scan after git push or gh pr create.
When to use it, and when not
Use it on code you own and trust: changes on a feature branch, a whole-repository scan before a release, patches for what it confirmed. For code you do not trust, the authors point to sandbox-runtime for the whole session.
A scan runs unattended for minutes to tens of minutes and spends a lot of tokens, so it is no pre-commit check. It needs the Workflow tool and Python 3.9+, and stops rather than improvising when the workflow is missing.
How to run it
We do not copy this skill: the plugin's own LICENSE is proprietary to Anthropic — all rights reserved, no redistribution — even though the repository root is Apache-2.0. Install it from Anthropic's repository with the button to the source. It is a Claude Code plugin from the official marketplace and works only in Claude Code: /plugin install claude-security@claude-plugins-official, then /reload-plugins; if the marketplace is missing, /plugin marketplace add anthropics/claude-plugins-official first. The plugin brings the skill with a prefix, /claude-security:claude-security, together with its hooks and subagents.
- Claude Code:
/claude-security, or ask for the skill by name. - Codex and Cursor: not supported. A scan runs only through Claude Code's Workflow tool, and the plugin's licence allows its use solely with Claude Code or other Anthropic products.
Example: /claude-security Scan my branch's changes against main at high effort; I know it may take a while and use a lot of tokens.
How it differs from similar skills
Security audit for codebases and services is the closest: hunters and verifiers over a coverage ledger, but it writes reports outside the repository and runs target code only in an OS sandbox. Here a scan only reads the code, while "Suggest patches" runs the project's own tests in a scratch clone, under your permissions and with no isolation of its own. Security review and Security review, Sentry are single-agent reviews of a diff or of the code you name. Repository threat model ranks abuse paths without hunting for them. Only this one gives SARIF and verified patch files.
Install
This skill is not copied to the catalog: its licence does not allow it. Install it from the author's repository.
Connect through AnyModel API
This skill runs on whatever model your agent uses. Point Claude Code, Codex or Cursor at AnyModel to run it on Claude, GPT, Gemini and more with one key — pay per token, no subscription.
AnyModel