Intel

AIKIDO-2026-74644

@anthropic-ai/sandbox-runtime is vulnerable to Sandbox Escape

Pre-CVE
Found by Aikido Intel before public disclosure or CVE publication.
Published Jul 22, 2026
July 22, 2026

78

CVE Score

High

Sandbox Escape

Are you affected? Scan for Free

TL;DR

On Linux the secure Bubblewrap branch isolates the PID namespace and mounts a fresh /proc but relies on bubblewrap auto-creating a user namespace, which only happens when the parent process runs with a non-zero effective UID. When the sandbox is launched from a root parent (effective UID 0, such as a container running as the default root user), the sandboxed command keeps CAP_SYS_ADMIN in the initial user namespace and can remount the read-only root bind read-write, escaping the filesystem restrictions the sandbox is meant to enforce. The fix forces --unshare-user and --cap-drop ALL in the secure branch so all capabilities are dropped regardless of the parent's effective UID. This closes the sandbox escape path for root-parent deployments.

How do I fix this?

Upgrade the @anthropic-ai/sandbox-runtime library to the patch version.

Who does this affect?

You are affected if you are using a version that falls within the vulnerable range and run the Linux sandbox from a process with an effective UID of 0, for example the default root user inside a container.

Versions

0.0.1 - 0.0.65

Fixed in 0.0.66

Background info

@anthropic-ai/sandbox-runtime is vulnerable to Sandbox Escape in versions 0.0.1 - 0.0.65.

Get secure today,
quickly and for free.

Secure your code, cloud, and runtime in one central system. Connect a repo to discover what the reasoning agents find in your codebase.

No credit card required | Scan results in 32 seconds.

Trusted by 150k+ orgs
  • Revolut
  • n8n
  • Lovable
  • Deel
  • Prove
  • Omnea