Researchers reveal Anthropic's Claude Co-Work exploited a Linux flaw to escape its sandbox, echoing recent security concerns surrounding OpenAI's ChatGPT agent vulnerabilities.
Claude Co-Work escapes sandbox via CVE-2026-46331 (CVSS 7.8), accessing SSH keys and cloud credentials on host system. Local deployment at risk; newer versions default to cloud execution but Anthropic released no dedicated fix, requiring manual hardening (namespaces, seccomp).
Researchers show Anthropic's Claude Cowork can escape its sandbox
Jul 27, 2026
09:37 am
What's the storyAnthropic'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.
SharedRoot exploits CVE-2026-46331 in local Claude Cowork sessions to gain guest root and read or write files across the host Mac.
Anthropic's new agentic AI system Claude Cowork has barely launched when security researchers documented a critical…
New trend could see humans lose control over AI systems, Anthropic warns
Anthropic’s Claude Sonnet 4.5 shows some “situational awareness,” raising safety and performance concerns.
Anthropic's advanced AI, Claude Mythos, capable of discovering software flaws, was kept from public release due to misuse…
The narrative around Anthropic's Claude, initially perceived as the "ethical AI company," has recently shifted due to concerns…
