This gives practitioners building or securing autonomous agents concrete, battle-tested containment patterns (sandboxes, VM boundaries, egress controls) instead of relying on human approval prompts, which the data shows fail due to rubber-stamping.
An Anthropic engineering post detailing the containment architecture used to limit the 'blast radius' of agentic Claude deployments across claude.ai, Claude Code, and Cowork.
It breaks down three risk categories (user misuse, model misbehavior, external attacks), explains why human-in-the-loop approval prompts suffer from approval fatigue (93% approval rate observed), and describes environment-level defenses like sandboxes, VMs, filesystem boundaries, and egress controls used instead.
