共有:

Claude Security Plugin

Claude Code now scans for vulnerabilities.

A multi-agent vulnerability scanner called "Claude Security" that runs right in the terminal has launched in beta. Writing code and reviewing it for security are starting to converge inside the same session.

AI Navigate Editorial2026.07.246 min read

/claude-security finding: F1 verify patches/F1.patch git apply
01

Why Now

Finishing security review
while you still write the code

Until now, writing code and reviewing it for security were two separate jobs.

In July 2026, Anthropic published the beta "Claude Security" plugin on its official Claude Code documentation. It's a multi-agent vulnerability scanner you call directly from a terminal-based Claude Code session. Installing it takes one command, "/plugin install claude-security@claude-plugins-official", distributed from Anthropic's official marketplace, the anthropics/claude-plugins-official repository.

Claude Code already ships a "security guidance" plugin that reviews code as Claude writes it, in the same session. Claude Security is a different tool, positioned as an on-demand, deep-scan layer. Aimed at a whole repository, or narrowed to a branch's diff, a pull request, or a single commit, a team of Claude agents splits the work: mapping the architecture, building a threat model, hunting for vulnerabilities, and independently verifying each finding. That's the decisive difference from traditional static analysis (SAST) tools, which just match known patterns via regex or signatures. Claude Security reasons about data flow across files and business-logic authorization design, in context.

Traditional static analysis (SAST)Claude Security
Matches known patterns (regex, signatures)Contextual reasoning via architecture understanding and data-flow tracing
High false-positive rate, needs manual triageAn independent verifier agent scrutinizes each finding before reporting
Flags issues only, no fixGenerates reviewed patches into a patches/ folder
Runs automatically in a CI pipelineRuns on demand inside a Claude Code session (paid plan)

Code an AI wrote,
an AI now second-guesses.


02

How It Works

A four-stage multi-agent pipeline

Not pattern matching — reasoning like a security researcher.

Map components and dependencies Map assets and attack surface Read code, hunt for issues A separate agent verifies 01 MAP 02 THREAT MODEL 03 HUNT 04 VERIFY Only confirmed findings land in CLAUDE-SECURITY-RESULTS.md
FIG. From mapping the architecture to independent verification, multiple agents split the work
v2.1.154
Minimum Claude Code version required
3.9.6+
Minimum python3 version required
3
Jobs in /claude-security (scan / diff / suggest patches)
01

Map the architecture

Read the repository and map its components and dependencies.

02

Build a threat model

Work out what assets matter and where the attack surface is.

03

Hunt for vulnerabilities

Read the actual code and surface anything suspicious.

04

Verify independently and report

A separate agent scrutinizes each finding, and only confirmed ones make the report.

A finding doesn't just sit there. Choosing "Suggest patches" assembles a patch candidate for each finding, tagged with an ID like F1. Here too, an agent independent of the one that wrote the patch runs the repository's tests if it has any, then checks three things before it will output a patch at all: does it address only that finding, does it introduce no new vulnerability, and does it leave other behavior unchanged. Patches land in a patches/ folder as files applied with git apply — nothing is ever applied automatically.

03

Who It's For

Useful for engineers
and PMs alike

Where it lands depends on your role.

Just the branch diff, before the PR

Scan "only this change" before merging. Narrow it to a pull request or a single commit hash, so fewer issues surface after review has already started.

Turn findings straight into patches

Have it draft git-apply-ready patches from the findings, one finding per pull request. When tests exist, they run before you see the patch, cutting the guesswork of where to start.

Scope down a large repository

Weighing file counts and relative cost, target just the API layer or the auth code. Useful for deciding, with a team's size in mind, where to spend security attention.


04

What's Next

Where to start

Running Claude Security requires Claude Code v2.1.154 or later on a paid plan, plus python3 (3.9.6 or later) on your PATH. The scan orchestrates its agents through what Anthropic calls dynamic workflows, so on a Pro plan you need to turn on "Dynamic workflows" in /config first. Once that's set, a solid first move is running /claude-security in a repository you own and picking "Scan changes" to look at just the branch you're currently working on. A full scan takes real time and tokens, so start narrow and get a feel for it.

05

Caveats

Don't over-trust it — it's one layer of defense

Anthropic's own documentation says scans are nondeterministic: two scans of the same code can surface different findings. That's exactly why it's built for repeated, ongoing runs rather than a single one-off scan.

Patches are designed so a human always reviews before applying — there is no auto-apply path at all. Anthropic also frames the plugin as an additional layer that sits in front of your existing static analysis, dependency (SCA) scanning, and CI checks, not a replacement for them. Given the beta status and paid-plan requirement, there's no need to force it into a hobby project. It's still no excuse to skip human review.