AI business agents at risk of hijacking with little user input

A major wave of AI-powered agents used in business settings appears vulnerable to hijacking with little to no user interaction, researchers say, potentially enabling data exfiltration, workflow manipulation and even impersonation of users. In demonstrations aligned with a high-profile industry conference, researchers showed how widely used AI assistants could be co-opted to bypass safeguards, undermine critical operations and establish long-term access to compromised environments. Vendors have responded with patches and safeguards, but experts warn that the rapid spread of agentic AI in enterprise makes governance, transparency and careful deployment imperative.

The vulnerabilities hinge on how autonomous AI agents operate across popular platforms that enterprises rely on for productivity, customer service and back-office tasks. Investigators demonstrated that attackers could influence what these agents do, alter their knowledge sources and, in some cases, cause them to act against the interests of the organizations they serve. One risk highlighted is the potential for memory persistence, meaning intruders could retain control or access well after an initial compromise. In...