Anthropic PBC announced the launch of Claude Security in public beta mode today to help cybersecurity teams scan their codebases for vulnerabilities and generate patches. 

Part of Claude Enterprise, the company’s subscription model aimed at large organizations, Security began as a research preview as Claude Code Security in February. Since its introduction, the company said hundreds of organizations have used it to discover and fix exploits in production code, including vulnerabilities existing tools had missed for years. 

The new security tool is powered by the company’s flagship artificial intelligence model, Opus 4.7, as a dedicated defensive product, part of a longer-term trend for the company as a cybersecurity angle.   

Anthropic recently launched Project Glasswing, debuting the use of the company’s Mythos model, which was not originally designed as a security model but has proved particularly adept at discovering vulnerabilities. Glasswing brings together numerous technology partners to secure the world’s production software. 

Following this initiative, the company said customers can now work with partners building on Claude’s security capabilities. This includes Anthropic’s technology partners, such as CrowdStrike Holdings Inc., Palo Alto Networks Inc., SentinelOne Inc., Trend Micro Inc.’s TrendAI and Wiz Inc., integrating Opus 4.7 into their cybersecurity platforms. 

Anthropic said that with these new capabilities, Claude can now reason over entire codebases by thinking similarly to the way a cybersecurity researcher does. Instead of looking for known patterns, it traces data flows, reads source code and examines interactions between code components and files to synthesize network effects. 

The model then verifies everything with a confidence rating before it reaches an analyst so that it can provide more trustworthy outputs. Anthropic also said the model will do its homework as it goes, explaining its reasoning along the way, its confidence factors, the likelihood that a vulnerability can be exploited, the triage factors and how effective a fix will be. 

From there, users can open a Claude Code session to apply the fix directly in context, with no more waiting days to go back and forth with security and engineering teams to get it tested and working. 

Anthropic said that after feedback from the research preview, the company added scheduled scans for ongoing coverage, the ability to dismiss findings with documented reasons (essentially notes for future reviewers to triage analysis), and comma-separated and Markdown exports for easy import into existing audit systems. 

Image: Anthropic