Experts say they were able to create a rogue agent in Google’s AI platform with just a single edit permission

One compromised agent could take over every other agent in that project, leading to chat logs access, and even data exfiltration.