Anthropic ships managed agents—and gates Mythos

Anthropic is trying to sell two opposite ideas at once: let companies hand more work to autonomous Claude agents, and keep its strongest Claude model away from the public because it may be too good at cyber offense. Both moves landed this week. (platform.claude.com) (euronews.com) The first move is Claude Managed Agents, a public beta that lets developers create an agent, give it tools, put it in a cloud environment, and run a live session without building all the plumbing themselves. Anthropic’s own docs break that into four pieces: agent, environment, session, and events. (platform.claude.com) In plain English, Anthropic is taking the part most companies usually build from scratch — the container, tool access, state, and long-running task loop — and turning it into a hosted product. Anthropic’s engineering team says the goal is a small set of interfaces that stays stable even as the underlying “harness” keeps changing. (anthropic.com) That “harness” is the scaffolding around a model: the reset rules, tool calls, memory handling, and stop conditions that keep an agent working for more than one prompt. Anthropic says those assumptions go stale as models improve, and it gives one example where a fix built for Claude Sonnet 4.5 became unnecessary on Claude Opus 4.5. (anthropic.com) At the same time, Anthropic moved Claude Cowork closer to standard enterprise software. On Anthropic’s product page, Cowork is now described as “ready for enterprise,” with admin controls for feature access, spending, and usage tracking across an organization. (claude.com) Cowork is the desktop version of this idea: instead of asking a chatbot one question at a time, a worker can hand Claude a multi-step task and come back later to a draft, spreadsheet, or report. Anthropic says Cowork can schedule tasks, organize files, build spreadsheets, prepare reports, and work across local files. (claude.com) The enterprise push is also about connectors, which are the pipes into the apps companies already use. The New Stack reports that Cowork’s new connectors include Google Drive, Google Calendar, Gmail, DocuSign, Apollo, Clay, Outreach, Similarweb, MSCI, LegalZoom, FactSet, WordPress, and Harvey. (thenewstack.io) Then comes the other half of the story: Anthropic says its new Claude Mythos Preview is its most powerful model yet, but it is not making it generally available. The company says the model is unusually effective at finding high-severity vulnerabilities in major operating systems and web browsers, which raises the risk of abuse by criminals and spies. (euronews.com) (indianexpress.com) Instead of a broad launch, Anthropic put Mythos inside Project Glasswing, a limited defensive program for a small group of partners. NBC News reports that more than 50 organizations, including Microsoft, Nvidia, and Cisco, are getting access, along with more than $100 million in usage credits, to find and fix weaknesses in widely used systems. (nbcnews.com) Anthropic’s pitch is that agents should become routine software for office work, while frontier cyber capability should stay fenced in. That is a very different product strategy from “ship the biggest model to everyone,” and regulators have publicly treated the caution around Mythos as evidence that at least one major lab is willing to slow down when the failure mode looks like real-world intrusion. (euronews.com)