Anthropic opens Claude Security beta

Anthropic has turned its security tooling into a product. On April 30, the company opened Claude Security in public beta for Claude Enterprise customers, after first testing it as “Claude Code Security” in a limited research preview. The pitch is pretty specific — let Claude inspect a codebase, decide which findings are real, draft fixes, and route work to the right people, all inside Claude rather than through a custom agent stack. That matters because security teams are drowning in alerts already, and AI is about to generate a lot more of them. (anthropic.com) ### What actually launched? Claude Security is a dedicated workflow inside Claude, not just a prompt template. Anthropic says Enterprise users can open it from the Claude sidebar or a dedicated security page, choose a repository, directory, or branch, and start a scan without wiring up APIs or building their own orchestration layer. The product page frames it as “scan to fix” — find issues, validate them, and propose patches for human review. (securityweek.com) ### Why not just use regular Claude? Because this is really about workflow, not raw model access. Plenty of companies can already point a strong model at code. The hard part is making that useful for a security team that needs evidence, prioritization, and a clean handoff into remediation. Anthropic is packaging those steps into something narrower and safer than “give a frontier model the w(securityweek.com)ot open-ended offensive capability. (anthropic.com) ### What model is underneath it? Anthropic says Claude Security runs on Claude Opus 4.7. That matters because the company is selling the tool on reasoning, not signature matching. The claim is that Claude can trace data flows across a codebase and catch bugs that simpler pattern-based scanners miss. In other words, it is trying to act less like a lint rule and more like a patient security researcher reading the whole system. (securityweek. ([anthropic.com)-counter-ai-powered-exploit-surge/)) ### Did Anthropic show any proof? Yes — and this is the number doing most of the work. Anthropic’s launch materials and follow-on coverage say the private preview found 500 production vulnerabilities in two months, including issues that had survived prior expert review. That does not prove every enterprise will get the same result. But it does show why vendors think this category is ready to move from demo to budget line. (pasqualepillitteri.it) ### Why launch this now? Because the company has been building toward “AI for defenders” for a while. Last year Anthropic argued that frontier models were becoming practically useful for vulnerability detection and remediation, not just theoretically interesting. Claude Security is the product version of that thesis — less a new scientific jump than a decision to commercialize the capabilities it has been tuning. (anthropic.com) ### So where does Mythos fit in? Mostly as the political backdrop. Anthropic has separately described Mythos Preview as a much more capable cyber model, with enough ability that it kept access tightly limited to approved defensive partners. This week, reports said the White House opposed Anthropic’s plan to broaden Mythos access to many more organizations. Claude Security does not use Mythos, but it launches into that same argument about how much cyber capability should be packaged and distributed. (crn.com) ### What’s the real catch? Success creates its own mess. If AI systems get better at finding subtle vulnerabilities, companies will not just get safer — they will also get flooded with more findings, more patch suggestions, and more decisions about what is urgent. A tool that can validate and route fixes is useful partly because the next bottleneck is not discovery. It is triage. (crn.com)opic is no longer just saying advanced models can help defenders. It is charging enterprises for a concrete version of that claim. The bigger fight now is not whether AI can find bugs. It is who gets access to the strongest versions, and under what guardrails. (anthropic.com)