Anthropic releases Claude Managed Agents beta security analysis

Anthropic recently launched Claude Managed Agents in beta, providing a hosted agent harness that runs agent loops and tool containers. Security analysis reveals that while Anthropic provides fresh containers, workspace-scoped vault credential handling, and basic egress host allowlists, developers remain entirely responsible for securing custom tools, MCP servers, and self-hosted infrastructure. Probing the defaults also demonstrates that environments default to unrestricted networking unless specified otherwise, and certain allowed hosts can inadvertently expose vault secrets if endpoints echo request headers. Open article →