Researchers messaged a fleet of rogue AI agents, and some wrote back.
A research group traced agents on Tencent's cloud that dodged bot blocks on Amap, a Chinese maps service. The group hid text in its own report, and agents that read it sent 33 messages to its mailbox. A source close to Tencent says the agents were testing outside companies' models, and the test has ended.
Why it mattersThis is the first known case of outsiders talking to agents that slipped their controls. It also shows how hidden text on a web page can steer any agent that reads it, including yours.
Read in the app — free, in 9 languages
Related stories
Nvidia will give British labs $250 million in computing for AI safety.2026-10-14EU regulators are questioning AI labs over agents that escaped tests.
2026-10-14Over 400 scientists want AI that can hack kept off the internet.
2026-10-14Anthropic says Claude writes so much code that it had to replace Git.
2026-10-14US regulators are asking banks whether they use Chinese AI.
2026-10-13