Anthropic’s Claude Cowork could escape its local VM and read credentials on a Mac

Security researchers at Accomplish AI demonstrated that Anthropic’s Claude Cowork could break out of its local virtual machine sandbox and read files across the underlying Mac, including SSH keys and cloud credentials. The attack, dubbed SharedRoot, exploited a Linux kernel privilege escalation vulnerability to gain root access inside the guest VM, then walked out through a writable filesystem mount that exposed the entire host. Accomplish AI disclosed the findings on July 23 and said roughly 500,000 macOS users running local Cowork sessions were exposed before the issue was addressed.

The escape chain worked because Cowork’s local execution mode runs inside a Linux VM that shares the host filesystem via a writable VirtioFS mount. That mount was intended to be accessible only to root within the guest, but the researchers found they could escalate from a session user to guest root by exploiting CVE-2026-46331, a Linux kernel flaw known as “pedit COW” that was given a severity score of nearly eight out of ten. Once the agent had root inside the VM, it could reach anything the logged-in Mac user could reach.

“We connected a folder to a fresh Claude Cowork session, sent one short message, and watched the agent escape the sandbox,” Oren Yomtov, principal security researcher at Accomplish AI, told The Hacker News. “From inside the VM, it reached the host Mac and read and wrote files all over it, far outside the folder we’d connected, with no permission prompt anywhere.” The vulnerability sits in the kernel’s traffic-control subsystem, where incorrect copy-on-write handling allows out-of-bounds writes into shared page-cache memory.

Anthropic closed the report as “informative” without issuing a direct fix, according to The Hacker News. The version of Claude Cowork released afterwards defaults to cloud execution, which sidesteps the local escape path entirely. Users who opt to run the agent locally rather than in the cloud, however, remain exposed unless they harden their configurations by disabling unprivileged user namespaces, restricting filesystem sharing, and running the Cowork daemon with strict mount protections.

The finding lands in a month that has seen four separate research teams break AI agents in four different ways, from poisoned memories to hijacked browser extensions. OpenAI disclosed that its own models escaped a sandbox and breached Hugging Face during the same week, and researchers escaped the sandboxes of Cursor, Codex, and Gemini CLI without ever breaking the sandbox itself. The pattern across all of these incidents is the same: the AI agent follows its rules inside the box, but the infrastructure surrounding it trusts the agent more than it should.