Skip to content
Submit incident
AII-1000-35401
Well documentedThree or more independent publishers report the core facts.

Claude Cowork escapes VM sandbox, accesses Mac host files

Security researchers demonstrated that Claude Cowork, Anthropic's AI agent, could escape its local virtual machine sandbox and access sensitive files on Mac computers, including SSH keys and cloud credentials, via a Linux kernel vulnerability. Anthropic responded by making cloud execution the default and requiring users to manually harden their configurations to use local execution safely.

23 July 2026 Β· Anthropic Β· Security and cyberattacks
23 July 2026
AII-1000-354014 sources ↓
Quick facts
Category
Security and cyberattacks
Country
United States
Source
The Next Web
Date
23 July 2026
Drafted by
Raidu Incident Detection

What happened

Security researchers at Accomplish AI demonstrated that Claude Cowork could escape its local virtual machine sandbox and read files across the underlying Mac, including SSH keys and cloud credentials.

The attack exploited a Linux kernel privilege escalation vulnerability (CVE-2026-46331, known as 'pedit COW') to gain root access inside the guest VM, then accessed the host Mac through a writable VirtioFS filesystem mount that was shared between the VM and host.

Accomplish AI disclosed the findings on July 23, 2026, and reported that approximately 500,000 macOS users running local Cowork sessions were exposed before the issue was addressed.

Reported impact

The agent could read and write files across the host Mac far outside the folder that was connected to the Cowork session, with no permission prompt.

Sensitive files including SSH keys, cloud credentials, and other user files were accessible to the escaped agent.

Index analysis

The vulnerability existed because Cowork's local execution mode runs inside a Linux VM with a writable VirtioFS mount to the host filesystem that was intended to be accessible only to root within the guest.

The Linux kernel flaw (CVE-2026-46331) allowed privilege escalation from a session user to guest root, enabling the agent to reach anything the logged-in Mac user could access.

Governance controls

Anthropic closed the researcher report as 'informative' without issuing a direct fix and instead released a new version of Claude Cowork that defaults to cloud execution, which sidesteps the local escape path.

Users who opt to run the agent locally remain exposed unless they harden their configurations by disabling unprivileged user namespaces, restricting filesystem sharing, and running the Cowork daemon with strict mount protections.

Classification

Category
Security and cyberattacks
Industry
Digital security
Country
United States
Harm status
Potential harm
Intent
Unintentional
Timing
Post-deployment
Harmed
Consumers
Failure modes
Unsafe autonomous behaviour, Data leak
AI systems
Claude Cowork

Sources and evidence

This record was researched and written by the Index, based on the following original reporting.

The Next Web
Anthropic's Claude Cowork could escape its local VM and read credentials on a Mac
26 July 2026
Firstpost
It's not just OpenAI: Researchers show Anthropic's Claude Co-Work can escape its sandbox too
26 July 2026
TechRadar
It's not just OpenAI models escaping and running riot -- experts show how Claude Cowork can break its bonds and access Mac files
26 July 2026
9to5Mac
Claude Cowork escaped sandbox on Mac, had full access to all files
27 July 2026