Anthropic has begun rolling out Claude Security, its multi-agent vulnerability scanner, as an installable plugin for Claude Code, the company’s terminal-based coding agent. The plugin entered public beta the week of July 21, 2026, according to Anthropic’s own product page, extending a tool that previously lived only inside the company’s managed Anthropic enterprise dashboard.
How the scanner works
Unlike simple pattern-matching linters, Claude Security is built to trace vulnerabilities that span multiple files. According to Anthropic, the plugin works in three stages: it first scans code in parallel while tracking context across the codebase, then runs an “adversarial verification pass” in which a separate reviewing agent challenges the initial findings, and finally proposes patches that a developer must review and approve before anything is applied.
The company says the tool targets “high-severity vulnerabilities including memory corruption, injection flaws, authentication bypasses, and complex logic errors” — the kind of context-dependent bugs that traditional security scanners often miss. Anthropic describes the underlying models as the same ones it uses to audit its own codebase.
Beta requirements
The plugin is distinct from Anthropic’s earlier, lighter “security-guidance” plugin, which runs continuously in the background during a coding session using simpler pattern checks. Claude Security is instead meant for deliberate, on-demand audits — either a full repository scan or a review of pending changes before a commit — and requires a paid Claude Code plan plus a local Python and Git setup. Organization administrators can turn it on for their teams from the Claude Code admin console.
The release lands as AI coding tools face growing scrutiny over the risks they can introduce as well as fix: Claude Cowork, Anthropic’s other agentic product, was found to have its own sandbox-escape flaw earlier this month. Anthropic is positioning Claude Security as evidence that the same agentic capability can also be turned toward hardening code before it ships.