Anthropic launched Claude Security in public beta on May 1, 2026, making it available at no additional charge to Enterprise tier subscribers. The product is built on Claude Opus 4.7, which reached general availability in April and carries a 1.2M-token context window, a specification confirmed by prior Anthropic documentation.
The capability Anthropic is selling here is autonomous end-to-end remediation: find the vulnerability, generate reproduction steps, and produce a working code fix, all in one session. According to Anthropic, the tool uses a specialized security scaffold to trace data flows and component interactions rather than relying on pattern matching. That architectural claim appears only in Anthropic’s own announcement; no independent source has confirmed how the security layer actually works. Anthropic says the product includes native exports to CSV and Markdown, with webhook integrations for Jira and Slack.
On verification: Anthropic cites internal red-teaming as its evaluation method. No external benchmark has been disclosed, and no Epoch AI evaluation is scheduled or complete. The 1.2M-token context window is the one figure in this brief with independent corroboration, everything else rests on Anthropic’s announcement alone.
Why does this matter to security teams? The product category itself isn’t new. Automated vulnerability remediation tools from SentinelOne, Veracode, and others have existed for years. What’s different here is the integration of a frontier reasoning model, specifically one that, in the restricted Mythos context, has already demonstrated the ability to outpace human patch cycles. Claude Security represents Anthropic’s attempt to bring some version of that capability into a commercially licensed, enterprise-accessible product.
That distinction matters because it changes the risk calculus for security teams evaluating whether to adopt the tool. An LLM that can reason about code structure across a 1.2M-token context has a different threat model than a signature-based scanner. If the model can find previously unknown vulnerabilities, it can also generate novel exploits. Anthropic doesn’t address that surface in what it has disclosed. That’s the practical question any DevSecOps team should ask before deploying it on production codebases.
Context: Anthropic has been building toward enterprise security tooling since the Mythos disclosures. The Mythos governance story established that Anthropic had a restricted-access model with real-world exploit-finding capability. Claude Security appears to be a commercially sanitized version of that capability, with Enterprise tier access controls instead of government-agency gate-keeping. Whether the security scaffold meaningfully constrains the model’s offensive potential is unverified.
What to watch: Three things. First, whether Anthropic publishes any external evaluation results during the beta period, the absence of disclosed benchmarks is notable for a security product. Second, how enterprise security vendors respond, since Claude Security enters a market with established players who have audit trails and compliance certifications that a beta LLM tool does not. Third, whether compliance officers at regulated enterprises treat this as an approved tool or a new risk surface requiring its own assessment before deployment.
The launch is real. The capability claims are Anthropic’s. Security teams should evaluate it accordingly, with the same structured threat-modeling process they’d apply to any agentic tool with code-execution proximity.