Anthropic expands Claude security beta

Claude is turning two knobs at once. One knob is scientific capability — especially in biology. The other is defensive security tooling for companies. Anthropic’s latest move matters because those two things usually pull against each other: a model that gets better at hard research and hard code analysis can help legitimate users a lot, but it also raises the stakes around misuse. ### What actually changed? Anthropic put Claude Security into public beta for Claude Enterprise customers. The product scans a company’s codebase, validates findings, suggests patches, and lets teams export or route results into tools like Slack and Jira. Every patch still needs human review and approval. At almost the same time, Anthropic published new biology results showing its latest models are getting meaningfully better at open-ended bioinformatics work. In its new BioMysteryBench evaluation, Claude Mythos Preview solved 30% of human-difficult problems that had previously stumped a panel of experts, while newer Claude generations generally performed on par with human experts. ### What is Claude Security, exactly? Basically, it’s Anthropic’s pitch that AI can act less like a noisy lint tool and more like a real security researcher. Traditional scanners mostly look for known patterns. Claude Security is supposed to reason across files, trace data flows, understand business logic, and catch context-dependent bugs that rule-based tools often miss. That matters because security teams already drown in findings. A tool that finds subtle vulnerabilities is useful only if it also filters junk. Anthropic says Claude runs an adversarial verification pass on its own findings before surfacing them, then attaches proposed fixes and confidence signals so analysts can triage faster. ### Why is Anthropic pushing security so hard? Because Claude’s cyber skills have already crossed from benchmark bragging into real-world output. In March, Anthropic said Claude Opus 4.6 found 22 Firefox vulnerabilities in two weeks, including 14 that Mozilla rated high severity. Anthropic also said that was more vulnerabilities in a month than Firefox had seen reported from any source in any single month of 2025. So the company’s strategy is pretty clear — keep the strongest capabilities pointed at defenders first, and wrap them in tighter access controls. You can see the same pattern in Opus 4.7, which shipped with automatic blocking for prohibited or high-risk cyber requests and a Cyber Verification Program for legitimate security professionals doing red-teaming or penetration testing. ### What does the biology result mean? It does not mean Claude became a lab scientist overnight. BioMysteryBench is a bioinformatics benchmark — so the work is closer to analyzing datasets, forming hypotheses, and reasoning through messy biological evidence than physically running wet-lab experiments. But the result is still notable. Anthropic says the newest models not only matched human experts overall, but sometimes solved problems the expert panel could not, and used different strategies from humans. That’s the interesting part. When a model starts finding alternate routes through hard scientific problems, you stop thinking of it as just a fast summarizer. ### Why pair biology gains with security controls? Because stronger models create a governance problem, not just a product opportunity. Anthropic has been saying for months that frontier models are becoming more capable in both cyber and biology, and that deployment has to get more selective as those capabilities improve. Its own security product page now puts Claude Security in public beta for Enterprise customers only, with Team and Max access still “coming soon.” That limited rollout is the tell. Anthropic wants enterprise revenue from high-value use cases, but it also wants to show that access can be narrowed, monitored, and reviewed while the underlying models keep getting stronger. ### So what’s the bottom line? Anthropic is no longer just selling a chatbot. It’s selling a controlled high-capability system for two expensive, high-stakes kinds of work — securing code and accelerating scientific analysis. The promise is huge. The catch is obvious. The better Claude gets at reasoning through vulnerabilities and biology, the more Anthropic has to prove it can decide who gets that power, and under what rules.