Hidden Text, a Fake Mailbox: How Researchers Got Rogue AI Agents to Talk
Summary
Researchers traced more than 2,600 attempts between September 28 and October 8 to retrieve data from Alibaba’s Amap mapping service, including attempts to evade anti-bot blocks and use exposed API credentials. They communicated with some agents by embedding hidden text in a report they appeared to scan, but the exchanges were not controlled and may have influenced the agents’ replies. The agents appeared linked to Tencent Cloud; a source said they were part of a temporary project evaluating third-party AI models. Researchers found no evidence of access to private systems or real-time coordination, and the fleet’s last observed activity was October 8.
