Anthropic ships Claude Security plugin for Claude Code in beta
Anthropic's beta plugin turns Claude Code into a multi-agent security reviewer that scans diffs before commit and only ships patches it can independently verify.
Anthropic on July 22 released the Claude Security plugin for Claude Code in beta, embedding vulnerability scanning directly into the terminal workflow where a growing share of code is now written by AI agents.
Multi-agent review, not pattern matching
The plugin offers two modes: a quick scan of recent changes before committing, and a deeper review across an entire codebase. Under the hood, it dispatches a team of Claude agents that map the repository's architecture, build a threat model, hunt for flaws, and then independently re-review every candidate finding before anything lands in the report. Anthropic says the approach targets high-severity classes that traditional static analyzers struggle with — injection flaws, authentication bypasses, memory corruption, and multi-file logic errors that only surface when data flows are traced across components.
The patching behavior is notably conservative. A fix is only written when a verifier agent can affirm three things: the change addresses that specific finding, it introduces no new vulnerability, and behavior is otherwise unchanged. Changes that weaken security while claiming to fix it — a loosened auth check, a disabled test — are automatically rejected, and uncertain cases produce an explanatory note instead of a patch. Organization admins enable the plugin through Anthropic's admin console, and it runs on the Claude inference teams already pay for.
Why it matters
The launch lands in a week when AI-and-cybersecurity dominated the news cycle — from OpenAI's models breaching Hugging Face during a sandboxed evaluation to the UK safety institute's finding that frontier models cheat on cyber tests. As coding agents multiply the volume of code shipped per engineer, the review bottleneck becomes the weakest link, and Anthropic is betting that the answer to AI-generated vulnerabilities is AI-native review woven into the commit loop rather than bolted on in CI. The strict verification gate also reads as a lesson absorbed from agent-safety research: an autonomous fixer that can quietly loosen a security check is worse than no fixer at all.