Anthropic previews Claude Code Security

Anthropic opened a limited research preview of Claude Code Security, a capability built into Claude Code that scans codebases for vulnerabilities and proposes patches for human review. The company pitched it as a departure from conventional static-analysis tools: rather than matching known vulnerability patterns, it said the system “reads and reasons about your code the way a human security researcher would,” tracing how components interact and how data moves through an application to catch flaws that rule-based scanners miss.

Anthropic built in several checks against the tool acting on its own judgement. Findings go through multi-stage verification and receive severity ratings before being surfaced, and any suggested fix requires explicit human approval before it can be applied — the company did not offer an option to auto-merge patches. The preview launched first for Enterprise and Team customers, with expedited access offered to maintainers of open-source repositories, a group with limited budget for dedicated security review.

The release sat alongside a wider push by frontier labs to position their coding agents as security tools rather than only productivity tools — a framing that cuts both ways, since the same reasoning capability that lets a model find a vulnerability can in principle help someone exploit one. Anthropic did not publish benchmark figures for the preview’s detection rate or false-positive rate at launch, leaving its practical usefulness against established scanners untested in public.