Anthropic Adds Multi-Agent Security Scans to Claude Code

For practitioners, multi-agent code scanning introduces an additional review layer between AI-assisted implementation and conventional application-security controls. Industry context: tools that combine codebase context, change-level analysis, and human approval can reduce time spent triaging candidate issues, but their findings still require verification against the application's threat model, deployment configuration, and business logic.

Anthropic has made the Claude Security plugin available in beta for Claude Code. According to Anthropic's documentation, the plugin can scan an entire codebase or a narrower target such as a branch diff, pull request diff, or individual commit. It runs within a Claude Code session and counts against the user's plan usage limits.

Anthropic documents a multi-agent workflow in which Claude agents map repository architecture, build a threat model, search for vulnerabilities, and independently review each finding before a report is written. The company describes the output as findings that developers can turn into suggested patches, with the developer reviewing and applying patches.

Anthropic's product page describes capabilities including cross-file data-flow tracing, adversarial verification of findings, and suggested patches. Those are vendor-described capabilities rather than independently measured detection or false-positive results.

The plugin is distinct from other Claude Code review functions, according to Anthropic's documentation:

• •Inline review examines code while Claude writes it.
• •/security-review performs a single pass over a branch.
• •/review examines pull requests.
• •Claude Security is intended for broader multi-agent repository or change scans.

Anthropic lists Claude Code version 2.1.154 or later, a paid plan with dynamic workflows enabled where applicable, Python 3.9.6 or later, and Linux, macOS, or Windows as prerequisites. Git is required for change scans and patch workflows, although Anthropic states that full scans can run in a directory without version control.

According to Anthropic, the plugin can reach repositories on GitLab and Bitbucket and repositories on networks that do not permit inbound connections. Anthropic separately offers a managed Claude Security product for Enterprise customers that monitors repositories; its product page identifies that service as being in public beta.

For practitioners, comparable AI-assisted security products are best evaluated as a supplement to, rather than a replacement for, static analysis, dependency scanning, secret detection, code review, and penetration testing. Teams can benchmark the plugin on previously remediated vulnerabilities and representative pull requests, then measure precision, recall, review time, and the quality of proposed patches.

For practitioners, keeping patch approval with a human reviewer is particularly important where a remediation changes authorization flows, input validation, cryptographic handling, or infrastructure-facing code. Anthropic states that every proposed patch requires review and approval, preserving that control point in the documented workflow.