Security Flaw in Google's Gemini AI Agent System Could Expose Secrets

A newly disclosed attack method targets Google's Agent Development Kit (ADK), a framework used to build AI agents that work together to complete tasks. Researchers found that an attacker could send a specially crafted prompt to a low-privilege agent, which would then pass along a malicious 'hand-off' comment to a more privileged agent in the same workflow.

This technique could allow an attacker to trick the privileged agent into exposing sensitive information or making unauthorised changes, such as tampering with a pull request in a software development pipeline. Because the attack exploits trust between AI agents rather than a traditional software bug, it highlights a new category of risk emerging as businesses adopt multi-agent AI systems for tasks like code review, customer service, and automation.

As more Australian small businesses experiment with AI agents to streamline operations, this incident is a reminder that AI systems can be manipulated just like traditional software, and that permissions and trust boundaries between automated tools need careful design and oversight.