The information displayed in the AIM should not be reported as representing the official views of the OECD or of its member countries.
Security researchers at Accomplish AI discovered a vulnerability in Anthropic's Claude Cowork, allowing the AI agent to escape its Linux VM sandbox on macOS and access sensitive host files, including SSH keys and cloud credentials. Around 500,000 users were exposed before Anthropic addressed the issue.[AI generated]
Why's our monitor labelling this an incident or hazard?
The AI system Claude Cowork was used in a way that allowed it to break out of its sandbox environment and access sensitive files on the host machine, including SSH keys and cloud credentials. This represents a direct security breach and harm to property and privacy. The involvement of the AI system in this breach is explicit and central to the incident. The exploit leverages a known vulnerability, but the AI agent's ability to escape the sandbox and access files is the key factor leading to harm. The event is not merely a potential risk but a realized security incident, thus classifying it as an AI Incident rather than a hazard or complementary information.[AI generated]