Anthropic Shipped an AI Code Security Scanner Into Claude Code

Anthropic Shipped an AI Code Security Scanner Into Claude Code — Here's What It Actually Catches

If you're using Claude Code, there's a new plugin worth turning on: Claude Security, now in public beta. It's not another regex-based linter bolted onto your pipeline — it's a multi-agent AI pipeline that reasons about your codebase the way a human security researcher would.

How it works:

1. Architecture mapping — builds a model of how your files connect, where data enters, where it ends up
2. Threat identification — traces data flow across files to catch injection points, unsafe deserialization, broken auth, multi-file logic errors
3. Adversarial verification — challenges its own findings before surfacing them (this is the part that actually cuts false-positive fatigue)
4. Patch suggestion — proposes a style-matched fix; nothing auto-commits, you review and approve every change

The bug class that sold me on this: an admin API where a permissions check lives in one file, but the DB write it's supposed to gate happens three function calls later in a totally different module. File-by-file SAST tools see two clean files. Claude Security's architecture-mapping step traces the actual call path and flags that the write is never really gated. Classic broken access control — one of the hardest bug classes for automated tools to catch.

Getting started:

Enable it org-wide from the admin console:
claude.ai/admin-settings/claude-code

Then run a review on your current session anytime:
/security-review

Or wire it into CI/PR reviews with Anthropic's open-source GitHub Action:
anthropics/claude-code-security-review

It scopes to changed files by default (cheaper, faster), or you can run a full-repo scan for deeper coverage — just know that costs more in tokens since it may shift to pricier models under the hood.

Vulnerability classes it targets: memory corruption, SQL/command injection, auth bypasses, insecure deserialization (including Python pickle risks), DOM-based XSS (dangerouslySetInnerHTML, .innerHTML), command injection (child_process.exec(), os.system()), and unsafe dynamic execution (eval(), new Function()).

I wrote a full breakdown — rollout timeline, setup, vulnerability classes, and best practices for not drowning in false positives:

https://www.xpert4cyber.com/2026/07/claude-security-ai-vulnerability-scanner.html

Anyone else running this against a real repo yet? Curious how the confidence scoring compares to Semgrep or CodeQL in practice.