Anthropic to broaden Mythos-class model access and ships 28 enterprise security integrations
Anthropic is moving its most capable — and most restricted — model class, internally referred to as 'Mythos', toward broader availability after months of staged red-teaming. The headline number from UK AI Security Institute testing is stark: Mythos achieved full corporate-network takeover in 6 of 10 attempts in a controlled environment, double the 3/10 rate that Politico reports for GPT-5.5. Anthropic is reportedly considering exposing Mythos-class capability through Claude Code, where the agent already has tool execution and filesystem access, which would significantly raise the ceiling of what coding agents can autonomously do.
In parallel, Anthropic is locking down the enterprise flank. The company announced 28 new security and governance integrations spanning the entire modern security stack — endpoint (CrowdStrike), network and SASE (Palo Alto Networks, Zscaler, Cloudflare, Fortinet, Netskope), identity (Okta), cloud-security posture (Wiz), and Microsoft's broader portfolio. The pitch is that enterprises adopting Claude in regulated environments can route activity, alerts, and policy enforcement through tools they already operate, rather than treating Claude as an opaque black box.
A third Anthropic thread this week: the company is planning a personal AI Fluency scorecard inside Claude, building on its February 2026 study of roughly 9,830 conversations that profiled how individuals use the model. The scorecard would surface a user's prompting patterns and effective-use signals back to them — part product feature, part data-driven case for upskilling.
The combination is editorially significant. Anthropic is simultaneously arguing that its most dangerous capabilities can be released responsibly (Mythos), that its enterprise plumbing is now mature (28 integrations), and that its users can be measurably 'fluent' on the platform. Security teams will watch the Mythos gating closely; the 6/10 corp-network-takeover figure is the kind of number that draws regulator attention, and the move from research-only to Claude-Code-adjacent surfaces is exactly the transition red-teamers warn about.