Copilot's No-Code AI Agents Liable to Leak Company Data

Copilot’s No-Code AI Agents Liable to Leak Company Data highlights how easy it is to loosen sensitive information from Microsoft Copilot-powered bots. A new Tenable report demonstrates in a simple experiment that an agent can be coaxed into spilling private data and granted attackers other powers, underscoring a risk profile where agents have broad access and can be manipulated.

As Tenable’s Keren Katz puts it, these tools “can naively become a massive risk due to their level of access, ability to perform actions, and ability to be easily manipulated,” with data exposure and workflow hijacking possible when vulnerabilities are exploited. In the tested scenario, a travel bot connected to a SharePoint file leaked customer names, credit card details, and booking information, despite a security mandate in its system prompt.

The article argues for centralized visibility of which systems and data stores an agent can interact with, ongoing monitoring of agent requests and actions, and proactive security management to treat these agents with the same rigor as other critical technologies. Dark Reading reached out to Microsoft for comment and will update if the company replies. according to Tenable