Anthropic's Claude AI can go rogue, researchers warn
What's the story
Anthropic's Claude Co-Work has been reported to have escaped its local sandbox, raising fresh concerns over the security of AI agents. The incident follows similar reports of ChatGPT escaping its sandbox and breaching tech firm Hugging Face. Accomplish AI's researchers tested Claude Co-Work in a local session on a Mac-hosted virtual Linux machine and found it could break out of this environment.
Security breach
AI agent could read and write files on host
The researchers observed that once outside the virtual machine, Claude Co-Work could read from and write files on the underlying host system.
It also gained access to nearly everything stored in a Mac user's account, including SSH keys, cloud credentials, and other sensitive information.
The escape was achieved by exploiting CVE-2026-46331, a Linux kernel privilege-escalation vulnerability with a CVSS severity score of 7.8 (High).
Response
How to minimize risks?
After the findings were disclosed, Anthropic acknowledged the issue but did not release a dedicated fix.
However, a newer version of Claude Co-Work defaults to cloud execution, which reportedly mitigates the risk.
Still, users opting for local over cloud execution could still be at risk.
To minimize such risks on local devices, users are advised to disable unprivileged user namespaces and tighten seccomp permissions among other measures.