Anthropic launches Claude Security beta to defend enterprise AI deployments

Security teams are getting a new kind of AI tool — one built for defense, not just coding help. Anthropic has moved Claude Security into public beta, opening it to Claude Enterprise customers after a smaller research preview earlier this year. The pitch is pretty direct: AI models are getting good enough at finding bugs that companies need equally capable systems to hunt those bugs first and help fix them. That makes this less like a shiny add-on and more like early infrastructure for the agent era. (claude.com) ### What is Claude Security? Claude Security is Anthropic’s code-scanning and remediation product for enterprise security teams. It looks through codebases for vulnerabilities, explains the issue in context, and suggests patches for human review. Anthropic first showed it in February 2026 under the name Claude Code Security as a limited research preview, then widened access on April 30 by rebranding it as Claude Security and putting it into public beta. (anthropic.com) ### Who can use it now? Right now, the product is available to Claude Enterprise customers. Anthropic’s own launch post frames it as something organizations can start using inside the Claude environment rather than a custom security stack that needs heavy setup first. A few outside writeups also note partner integrations and audit workflow support, which matters because security teams usually need findings t(anthropic.com)claude.com) ### What does it actually do? The core workflow is simple. A team points Claude Security at a repository — or just a specific directory — and the model scans for flaws, reasons through code paths, and proposes targeted fixes. Anthropic says the beta includes scheduled scans, targeted scans, and improved tracking for triaged findings. The interesting part is that this is meant to catch issues that pattern-matchi(claude.com)ather than just match known signatures. (claude.com) ### Why is Anthropic launching this now? Because the company thinks the offense side is about to get much faster. In the beta announcement, Anthropic says current models are already effective at finding flaws in software code, and that the next generation will be better at autonomously exploiting them. That warning lines up with Anthropic’s broader cyber work this spring, including its Mythos preview, which the(claude.com)pabilities and tightly controlled access. Basically, Anthropic is saying the exploit loop is speeding up — so defense has to automate too. (claude.com) ### Why does Opus 4.7 matter? Claude Security is built on Claude Opus 4.7, which Anthropic describes as its most powerful generally available model. That matters because vulnerability discovery is not just a search problem — it needs long-context reasoning across repositories, dependencies, and data flows. If the model is better at understanding how a system behaves, it should be better at spotting the weird ed(claude.com)roduct in beta, which tells you the company sees this as useful but not fully settled. (anthropic.com) ### Is this just another code assistant? Not really. A code assistant helps developers write features faster. Claude Security is aimed at the opposite side of the workflow — finding the dangerous parts before an attacker or a reckless internal agent does. The overlap is important, though. The same improvements that make AI better at writing and modifying software also make AI better at probing software for weaknesses. So (anthropic.com)ools helping teams ship agentic products are also expanding the attack surface those teams need to defend. (claude.com) ### What should companies take from this? The big takeaway is that AI security is shifting from policy talk to operational tooling. If a company is deploying internal agents, exposing model-connected workflows, or letting AI touch production code, it probably needs to think about continuous vulnerability review in a different way. Claude Security will not replace human security engineers. But it does look like (claude.com)— AI systems that can inspect, reason, and patch at machine speed while humans stay in the approval loop. (claude.com) ### Bottom line Anthropic is not just selling another model feature. It is making a bet that enterprise AI security will become its own product category — because the same models that help build software are getting close to being able to break it, too. (claude.com)