Source: Making Claude Code more secure and autonomous with sandboxing (Anthropic, 2025)

Anthropic engineering post on Claude Code's OS-level filesystem and network sandbox, which cut permission prompts by 84% internally, plus the cloud git-proxy design.

Contents

Summary

Argues that per-action approval prompts cause approval fatigue, and replaces them with pre-defined boundaries enforced by the OS. The agent can act freely inside the boundary; prompt-injected behaviour stays contained.

Key claims

  • Two isolation layers are needed together: filesystem (read/write limited to the working directory) and network (all traffic through an out-of-sandbox proxy that allows approved domains and asks for new ones). Without network isolation a compromised agent can exfiltrate SSH keys; without filesystem isolation it can escape and reach the network.
  • Implemented with Linux bubblewrap and macOS seatbelt, no container needed; covers spawned subprocesses; the runtime is open source; enabled via /sandbox.
  • Internal usage: sandboxing reduced permission prompts by 84%.
  • Threat model centres on prompt injection.
  • Claude Code on the web: each session runs in an isolated cloud sandbox; git credentials and signing keys stay outside; a git proxy authenticates the sandbox with a scoped credential, checks the operation (e.g. push only to the configured branch), then attaches the real token.

Feeds

Coding agents

Suggest changes →