Anthropic Beta Releases Claude Security Plugin for Code

Anthropic released the Claude Security plugin for Claude Code in beta this week, allowing users to perform multi-agent vulnerability scans of code repositories directly from within an existing Claude Code session. This new plugin aims to streamline the security auditing process by enabling users to identify vulnerabilities and then generate patch files for review and application. The tool offers flexibility, capable of conducting comprehensive scans across an entire codebase or focusing on specific changes just before a commit.

The plugin introduces a single command, `/claude-security`, which presents users with a menu of three distinct jobs. These include "Scan codebase" for the entire repository or a specified subset, "Scan changes" to analyze differences in branches, pull requests, or individual commits, and "Suggest patches" to convert identified vulnerabilities into `.patch` files. Installation is straightforward, requiring two commands from the Anthropic marketplace: `/plugin install claude-security@claude-plugins-official` followed by `/reload-plugins`. If the marketplace is not initially accessible, users can add it with `/plugin marketplace add anthropics/claude-plugins-official`.

The underlying scan pipeline is structured as a dynamic workflow orchestrated by a JavaScript script that distributes tasks among sub-agents. This process is divided into six distinct phases. The "Inventory" phase partitions the repository into manageable components, deciding which directories to scan or skip. "Threat model" assigns a modeler to each component to identify entry points, sinks, trust boundaries, and critical files. "Research" involves a dedicated researcher for each component and category, followed by a "Sweep" phase to fill any gaps identified by the matrix. The "Panel" phase employs a three-lens adversarial verification, and the "Adversarial" phase conducts maximum-effort red-teaming on any remaining vulnerabilities. The research phase specifically targets four categories: injection-and-input, authentication-and-authorization, and other security-related areas.