Anthropic's Security Review + cc-guard Hooks: Layered Defense

April 21, 2026 · 4 min read

Anthropic released claude-code-security-review — a GitHub Action that uses Claude to scan PRs for security vulnerabilities. It catches SQL injection, XSS, hardcoded secrets, and other OWASP Top 10 issues before code merges.

That's Layer 1: pre-merge defense. But what about the commands Claude Code runs right now, in your terminal, at 3am during an autonomous session?

That's where Layer 2: runtime hooks come in.

The Two Layers

Layer 1: Anthropic Security Review

PR submitted → Claude scans diff → blocks vulnerable code before merge

Layer 2: cc-guard Hooks

Command about to execute → hook inspects → blocks dangerous commands in real-time

Aspect	Security Review (Layer 1)	cc-guard Hooks (Layer 2)
When	At PR time	At execution time (real-time)
What it catches	Code vulnerabilities (SQLi, XSS, secrets in code)	Dangerous commands (rm -rf, force-push, credential deletion)
What it misses	Runtime actions, file deletion, git operations	Static code vulnerabilities, logic flaws
Speed	Minutes (full PR scan)	Milliseconds (pattern match)
Scope	Code changes only	All terminal commands + file writes
Token cost	Uses Claude API per review	Zero (bash scripts, no API calls)

What Layer 1 Doesn't Catch

The security review Action is excellent at finding code-level vulnerabilities. But 90 documented incidents show that most Claude Code disasters aren't code vulnerabilities — they're operational actions:

rm -rf ~/projects — 50GB permanently deleted (#49129)
git push --force origin main — untested code deployed at 3am
git add .env — API keys committed to public repo (#6527)
Auto-compact consumed entire token budget overnight (#51088)
Extended thinking burned 16M tokens in 25 minutes (#51092)

No PR review catches these — they happen in the terminal, between PRs, during autonomous sessions.

Setting Up Both Layers

Layer 1 — Add the official Action to your repo:

# .github/workflows/security-review.yml
name: Security Review
on: [pull_request]
jobs:
  review:
    runs-on: ubuntu-latest
    steps:
      - uses: anthropics/claude-code-security-review@v1
        with:
          anthropic_api_key: ${{ secrets.ANTHROPIC_API_KEY }}

Layer 2 — Install runtime hooks (30 seconds):

npx @gaebalai/cc-guard

That's it. Layer 1 reviews your PRs. Layer 2 blocks dangerous commands in real-time. They don't conflict — they complement.

Why Both Matter

A Stanford/MIT study found 14.3% of AI-generated code contains security vulnerabilities. Layer 1 catches those before merge. But the remaining 85.7% of "safe" code can still be deployed via git push --force to production at 3am — and Layer 2 is the only thing that stops it.

Neither layer alone is enough. Together, they provide defense-in-depth: one guards the code, the other guards the execution.

Get Both Layers

Layer 1: anthropics/claude-code-security-review (official)

Layer 2: npx @gaebalai/cc-guard

Test your Layer 2 — paste your settings.json, see what survives 10 real attacks