Anthropic publishes how it contains Claude across three products, using sandboxes and egress controls to cap what agents can reach
Synopsis
Anthropic has published the containment architectures behind three agentic products, claude.ai, Claude Code and Claude Cowork, shifting security from supervising what an agent does to limiting what it can reach. claude.ai runs code in a server-side gVisor container, Claude Code uses an OS-level sandbox that cut permission prompts by 84%, and Claude Cowork confines code execution to a virtual machine with its own kernel. The account says two of the incidents that taught the most both traveled through permitted egress paths, where the model layer had nothing anomalous to catch and only a deterministic boundary held.
Interpretation
Anthropic puts the center of agent security at the environment layer, using sandboxes, virtual machines and egress controls to set a hard boundary on what an agent can reach rather than supervising each step. Claude Code originally asked users to approve write, bash and network access turn by turn, but telemetry showed users approved roughly 93% of permission prompts, and approval fatigue made that oversight unreliable over time. After the OS-level sandbox shipped, permission prompts fell by 84%, and the runtime was open-sourced so the boundary can be audited.
The three products match isolation strength to whether a user can judge the risk of an agent's action: claude.ai executes code in a server-side gVisor container with a per-session filesystem, Claude Code runs on the user's machine with reads allowed, writes inside the workspace allowed and network denied by default, and Claude Cowork executes code in a VM with its own kernel and filesystem that mounts only the user-selected workspace and the .claude folder. Claude Cowork first ran the whole agent loop inside the VM, so any VM startup failure made the product unusable; the agent loop was later moved to the host while code execution stayed inside the VM. Claude Cowork keeps credentials in the host keychain and gives the VM a per-session scoped-down token that can be revoked independently of the user's.
Perspective
The account covers the containment designs and internal red-team exercises of three Anthropic agentic products, aimed at engineering and security teams building similar systems; stated next directions include classifiers at session startup and cross-platform agent identity standards.
The text states that model-layer defenses will never be 100% effective and that any probabilistic defense has a non-zero miss rate; Claude Mythos Preview was withheld in April 2026 because its blast radius was deemed too high, and the authors expect broader release of similarly capable models as defenders harden critical systems, though some risk will always remain.
