/security-review command: security audits in Claude Code

Plan eligibility

Availability depends on your Anthropic plan. As of 2026-05-12, the official help center article (Anthropic article 11932705) explicitly lists three compatible plans.

On the CI side, the GitHub Action needs a valid Claude API key stored as a repo secret. In practice that means an active API Console credit balance or a compatible plan that exposes a key usable outside the CLI.

Why /security-review matters

Anthropic's pitch is straightforward in the product blog post from 2025-08-06: turn a security review into a few-seconds gesture, baked into the coding workflow. Where a static scanner like Semgrep or Snyk applies a rule catalog (often calibrated against the OWASP Top 10), this AI security scanner leans on the model's reasoning to analyze code in context. The GitHub Action's README claims "deep semantic analysis" and "advanced false positive filtering", two promises you won't find in signature-based tools.

Three practical differences from a classic scanner:

1. The command reasons about the current diff. It sees the modified function and its callers, not just the flagged line.
2. It's language-agnostic. The Action's README spells out "language-agnostic", which saves you from installing a dedicated parser per stack.
3. It suggests a fix you can apply. Once a finding shows up, you can ask Claude to apply the patch right away, no manual copy-paste.

The flip side holds too: /security-review does not replace a static scanner. Anthropic's docs are firm on the word "complement", not "replace". More on that in the limits section.

Running the command locally

cd ~/projects/my-app
claude

/security-review

The primary help center source does not document any flag or argument for the command as of 2026-05-12. If Anthropic adds options later, /security-review --help inside the CLI should surface them.

Automating with the GitHub Action

The CI counterpart is the open-source repo anthropics/claude-code-security-review (ouvre un nouvel onglet), published under the MIT license. As of 2026-05-12, it sits at 4,573 stars, 432 forks and 69 open issues, latest push on February 11, 2026. The job fires on every pull request and posts findings inline as PR comments.

Minimal workflow

Here's the exact snippet the README ships with. It's fifteen lines long, scans the PR diff, and posts findings as comments.

name: Security Review

permissions:
pull-requests: write  # Needed for leaving PR comments
contents: read

on:
pull_request:

jobs:
security:
  runs-on: ubuntu-latest
  steps:
    - uses: actions/checkout@v4
      with:
        ref: ${{ github.event.pull_request.head.sha || github.sha }}
        fetch-depth: 2

    - uses: anthropics/claude-code-security-review@main
      with:
        comment-pr: true
        claude-api-key: ${{ secrets.CLAUDE_API_KEY }}

Two things to nail down during setup:

• CLAUDE_API_KEY secret: create it under Settings → Secrets and variables → Actions. The key needs access to both the Claude API and Claude Code.
• GITHUB_TOKEN permissions: pull-requests: write to post comments, contents: read to read the code. Nothing more.

Configuration options

The main options exposed by the Action are documented in the README as of 2026-05-12.

The Action is diff-aware: it only analyzes files modified by the PR, not the whole repo. This avoids the cost blowup you'd hit on any project past a few thousand lines.

The 5 official families detected

Anthropic's help center formalizes five vulnerability families. The Action's README lists more granular technical subtypes. The two lists don't contradict each other, the second is just finer-grained.

This five-family layout works as a reading grid. In practice you'll also see findings around weak crypto or insecure default configurations: Anthropic maps these under "insecure data handling" or "dependency vulnerabilities" depending on the case.

Limits and false positives

Four more limits to keep in mind:

1. Code is sent to the Anthropic API. Every call ships your diff to Anthropic's servers. For proprietary or NDA-bound code, vet your org's data policy and Anthropic's first.
2. Token cost. The command bills like any other API call. On a busy repo (dozens of PRs per day), watch consumption, especially if you wire the Action to every push.
3. Categories filtered out by default. To cut noise, the Action excludes Denial of Service, rate limiting, memory/CPU exhaustion, input validation without proven impact, and open redirects. If those angles matter to you, tweak the custom instructions.
4. No coverage guarantee. Anthropic says it plainly in the help center: "While automated security reviews help identify many common vulnerabilities, they should complement, not replace, your existing security practices." No precision or recall numbers are published.

Wiring the command into a DevSecOps workflow

/security-review shines when it sits inside a toolchain, not alone. Here's a defensive combo you'll often see on mature projects, paired with the security best practices page and the CI/CD and cybersecurity guide.

On the CLI side, you can also trigger the command automatically through a Claude Code PreToolUse hook on git commit. The mechanism is covered on the hooks page and stops you from forgetting the review before a push.

If an API key did leak through the cracks, the remediation playbook lives in Leaked API key recovery. To harden the MCP servers Claude calls during the review, see MCP security. For community agents dedicated to audit work (Security Reviewer, TDD Guide), the curated list is on Best security plugins.

On the threat model side, remember /security-review does not replace a manual pentest or a commercial SAST tuned to your stack. It catches a meaningful share of recurring mistakes (concatenated SQL, forgotten secrets, missing access controls) before they hit prod, which is already valuable.

Going further

Beyond the slash command, Anthropic ships a web product called "Claude Code Security". It's an Enterprise offering for continuous scanning of a full codebase, with a dashboard and grouped patch suggestions. As of 2026-05-12, this product is distinct from /security-review: the slash command stays available to all paid individual plans, the web product targets Enterprise teams. If you landed here looking for the web product, check Anthropic's official docs rather than this article.

Official links worth bookmarking :

• Automated security reviews in Claude Code (ouvre un nouvel onglet) (help center, primary source)
• Automate security reviews with Claude Code (ouvre un nouvel onglet) (product announcement, 2025-08-06)
• anthropics/claude-code-security-review (ouvre un nouvel onglet) (Action repo, MIT)

Next steps

• Run /security-review locally in your next Claude Code session, on a real diff
• Wire the GitHub Action into a secondary repo to calibrate cost and finding relevance before enforcing it in CI on critical repos
• Read the hooks page to automate the command before each commit
• Round out your defense with the practices laid out in security best practices and CI/CD and cybersecurity