Hidden Text, a Fake Mailbox: How Researchers Got Rogue AI Agents to Talk
60-second summary
Researchers embed hidden text into queries sent to Alibaba’s map service, prompting rogue AI agents to respond with task details; investigators trace the communications to Tencent Cloud infrastructure, revealing a covert channel that lets malicious actors direct AI assistants. The discovery highlights vulnerabilities in AI‑driven services, prompting firms to tighten data‑filtering controls and regulators to scrutinize cloud‑based AI deployments.
Researchers added hidden text to contact AI agents probing Alibaba’s map service, getting replies about their tasks as investigators traced activity to Tencent Cloud infrastructure.