Claude Security Plugin for Claude Code Now in Beta

I'll pull the product page and HN thread so the paragraphs stick to real details only—no invented numbers or dates.Pulling the docs and blog post for hard product mechanics.Anthropic has put the **Claude Security** plugin for **Claude Code** into beta. The plugin installs from the official marketplace with `/plugin install claude-security@claude-plugins-official` and exposes a single entry point, `/claude-security`, for full-repo scans, change-scoped scans, and patch suggestion. It is aimed at Claude Code users on a paid plan, with admins able to enable related security controls from the admin console. Early Hacker News coverage of the product page was thin: six points and one comment.

Technically, a scan is a multi-agent pipeline that runs inside a Claude Code session rather than against a hosted repo connector. Agents map architecture, build a threat model, hunt for issues, then send each candidate through an independent verification pass before anything lands in a report. Scope can be the whole tree, a focused area the plugin sizes for you, a branch or PR diff, or a single commit. Results write only into a timestamped `CLAUDE-SECURITY-<timestamp>/` directory: a human report (`CLAUDE-SECURITY-RESULTS.md`), machine-readable `jsonl`, a revision stamp tying findings to a commit (or `UNVERSIONED`), and optional `patches/` files. Patches are drafted in a scratch copy of the repo, reviewed by a second agent (including project tests when present), and applied only by the user—typically via `git apply`—never automatically. Prerequisites are Claude Code v2.1.154 or later (for dynamic workflows), Python 3.9.6+ as `python3` on PATH, and Git for change scans and patch flow.

For engineers and security builders, the useful split is local depth versus managed monitoring. The plugin reaches code the Enterprise **Claude Security** product may not: GitLab or Bitbucket hosts, air-gapped or inbound-restricted networks, and one-off trees that never get a cloud connector. It sits above lighter Claude Code surfaces—security guidance while Claude writes code, a one-shot `/security-review` on a branch, and PR **Code Review**—as the on-demand deep scan that still leaves remediation under human control. That matters when false positives burn analyst time and when “finding” without a reviewable patch just grows a backlog.

Market context is Anthropic stacking defender tools while model-side cyber capability rises. The broader **Claude Security** product (public beta for Claude Enterprise, previously previewed as Claude Code Security) runs scan-to-fix on connected repos with Opus 4.7, scheduled/targeted scans, dismissals with reasons, CSV/Markdown export, and webhooks into Slack or Jira. Partner channels put the same model class into CrowdStrike, Microsoft Security, Palo Alto Networks, SentinelOne, TrendAI, and Wiz, with services firms helping enterprises deploy. The Code plugin is the developer-machine half of that story: same scan/validate/patch idea, local session, plan usage limits, and no claim to replace static analysis or dependency scanners.

Practical takeaway: treat the plugin as a high-signal complement, not a CI replacement. Install only if you meet the version and Python floor; prefer auto mode so multi-agent workflows do not stall on permission prompts; scope large monorepos by subsystem and re-scan; always re-read patches and run your own tests before merge—especially when the patch note says the review ran without a test suite. Watch how often independent verification actually suppresses noise on your codebases, how Fable 5 cyber safeguards force downgrades mid-scan, and whether teams standardize on plugin reports in PR review versus the managed Enterprise path for continuous coverage.