Anthropic scales agents and funds security testing

Anthropic is trying to solve two different problems at once: companies want artificial intelligence agents that can actually run jobs for hours, and the internet is full of critical software that still ships with exploitable bugs. On April 8, 2026, it pushed on both fronts with a public beta for Claude Managed Agents and a new security effort called Project Glasswing. (claude.com) (anthropic.com) Claude Managed Agents is Anthropic’s hosted system for running an agent, which is software that can keep working through a task instead of answering one prompt and stopping. Anthropic says developers can define the task, tools, and guardrails, while Anthropic runs the infrastructure underneath. (platform.claude.com) (claude.com) The pitch is simple: most companies do not want to spend months building the plumbing around an agent before the agent does anything useful. Anthropic says production deployments usually need sandboxed code execution, checkpointing, credential management, scoped permissions, and tracing, and its service bundles those pieces together. (claude.com) Anthropic’s docs say these agents can read files, run shell commands, browse the web, and execute code inside managed containers. The same docs say they are built for long-running work that can last minutes or hours, with server-side event history and the ability to interrupt or steer a session while it is running. (platform.claude.com) Under the hood, Anthropic is separating the “brain” from the “hands.” Its engineering team says the model can change over time, but the outer interfaces stay stable, so customers do not have to rebuild everything each time Claude gets better or behaves differently. (anthropic.com) That sounds abstract, but the company gave a concrete example. Anthropic said an older harness tweak it used for Claude Sonnet 4.5 became unnecessary on Claude Opus 4.5, which is exactly the kind of brittle behavior companies hate when they are trying to ship software on a schedule. (anthropic.com) The bigger commercial bet is that agents move from demos to back-office systems only when someone else handles the boring, failure-prone parts. Anthropic is even using a beta header dated 2026-04-01 for the new application programming interface, which is a sign this is a fresh platform surface rather than a renamed chatbot feature. (platform.claude.com) On the same day, Anthropic announced Project Glasswing with a very different customer in mind: the people who maintain the software everyone else depends on. The launch group includes Amazon Web Services, Apple, Broadcom, Cisco, CrowdStrike, Google, JPMorganChase, the Linux Foundation, Microsoft, NVIDIA, and Palo Alto Networks. (anthropic.com) Anthropic says it formed Project Glasswing after seeing what it calls Claude Mythos 2 Preview, an unreleased model it says can outperform all but the most skilled humans at finding and exploiting software vulnerabilities. The company says that model has already found thousands of high-severity vulnerabilities, including some in every major operating system and web browser. (anthropic.com) The money behind that effort is not small. Anthropic says it is committing up to $100 million in usage credits for Mythos Preview and another $4 million in direct donations to open-source security organizations, while also extending access to more than 40 additional groups that build or maintain critical software infrastructure. (anthropic.com) Put together, these launches show Anthropic trying to become both the company that sells the agent runtime and the company that argues advanced models should be pointed at defense before attackers get the same tools. One product is aimed at enterprises that want agents in production, and the other is aimed at securing the operating systems, browsers, and open-source code those enterprises already run. (claude.com) (anthropic.com)