How we contain Claude across different products
我们如何对不同产品中的Claude进行隔离控制
Anthropic describes three mechanisms for containing Claude agent deployment risks across products: sandboxing or VMs, network egress controls, system-prompt and training constraints, and fine-grained permissions for MCP servers and third-party plugins.
Why it matters: An Anthropic engineer published the isolation controls that limit agent blast radius across claude.ai, Claude Code and Claude Cowork, addressing how reachable systems grow as models get more capable.