Anthropic has launched Claude Security in public beta, a suite of tools that utilizes AI agents to scan codebases for high-severity vulnerabilities and suggest patches. The new tools are designed to identify and patch complex software vulnerabilities through multi-agent reasoning.
Claude Security is available as a managed service for Enterprise users and as a plugin for Claude Code, Anthropic’s command-line interface for developers. The system moves beyond traditional static analysis security testing (SAST) by employing multi-agent workflows to scan code, build threat models, and conduct adversarial verification of findings.
The managed service is tailored for Claude Enterprise customers and offers continuous monitoring of repositories. It includes administrative controls, recurring scan scheduling, and webhook integrations for platforms such as Slack and Jira. In contrast, the Claude Security Plugin is aimed at developers working locally or within CI/CD pipelines and performs analysis without requiring inbound connections to repositories.
The system’s architecture relies on an orchestrated team of AI agents. Upon initiation of a scan, these agents analyze the codebase to understand its architecture and data flow. They then construct a specific threat model for the application and utilize reasoning-based approaches to hunt for vulnerabilities. This method allows for the identification of context-dependent issues, including authentication bypasses and complex logic errors.
Adversarial verification is a key component of the workflow. Each finding generated by the scanning agents undergoes a secondary review by a different agent, which attempts to disprove or challenge the results. Anthropic states that this step aims to reduce false positives, ensuring that verified issues are reported to security analysts or developers.
Claude Security focuses on detecting high-severity vulnerabilities, including memory corruption, injection flaws, and authentication failures. The system can read Git history and understand business logic, which enhances its effectiveness in finding multi-component patterns across various files in a repository. Once a vulnerability is identified and verified, the system generates a recommended patch that maintains the existing code’s structure and style.
