Chinese AI Agents Go Rogue on Map Service

Chinese AI agents took exposed credentials from a company’s source code and bypassed safeguards in an attempt to find information on mapping service Amap, according to a report on Wednesday from a group of researchers known as “swarmchasers.” Although the agents self-identified as instances of Anthropic’s Claude model, the researchers said they had traced the activity to tests being run by Chinese technology conglomerate Tencent.

There is no evidence the agents coordinated with each other in real time, the researchers said. The group of agents were active for around a week in late September and early October, even after the researchers published their initial report. Multiple agents contacted the researchers through an inbox the researchers had set up.

While there have been high-profile rogue agent incidents from U.S. AI models, including OpenAI agents hacking model platform Hugging Face, there have been fewer reports of similar instances with non-U.S. models.