Anthropic has published a detailed overview of their sandboxing techniques used across Claude products, including gVisor for Claude.ai, Seatbelt/Bubblewrap for Claude Code, and full VMs for Claude Cowork. The documentation reveals their multi-layered security approach using process sandboxes, VMs, and egress controls to prevent credential exfiltration and unauthorized access. The article also mentions past security vulnerabilities they've addressed and highlights their open-source srt (Sandbox Runtime) tool.
Background
As AI systems become more powerful and integrated into various products, ensuring their safe containment and preventing unauthorized access or data exfiltration has become a critical security concern. Sandboxing techniques are essential for isolating AI models from sensitive systems and data.
- Source
- Simon Willison
- Published
- May 31, 2026 at 05:36 AM
- Score
- 7.0 / 10