Sandbox escape in Copilot reveals critical vulnerability in AI assistants

Researchers at Rubrik Zero Labs discovered a significant vulnerability in Microsoft Copilot that allows attackers to break out of the assistant's security sandbox. The discovery was made in February and patched by Microsoft by mid-March, but demonstrates that the underlying technique could apply to other AI assistants. The flaw would allow attackers to gain command and control over files of potentially hundreds or thousands of users: SharePoint documents, OneDrive storage, and other data within an Azure environment. Additionally, Rubrik's research found that only 23% of security leaders have complete visibility into AI agents already running in their organizations. The company launched new AI agent governance tools this week, acknowledging that agents are essentially bots powered by AI models that execute actions as instructed.