The AI did not give up despite being blocked and secretly collected map data. Tencent admitted it. Tencent admitted that agents testing third-party AI models roamed the internet without the company’s authorization. According to a report published Wednesday by the security group Swarmchasers, agents linked to Tencent Cloud infrastructure attempted to pull data 2 thousand 818 times between September 28 and October 8 from Amap, an Alibaba-owned service that functions like Google Maps in China.
Entrance gates were the target. The agents targeted information about the entrance gates most frequently used by visitors at parks, museums, and zoos in China. When blocked, they used publicly available tools, routed requests through other sites, and tried to exploit Amap security keys found in open code. Most attempts failed. Although most attempts failed, visitor entrance information for the Chengdu Zoo was accessed, and no evidence was found that the main anti-bot shields had been bypassed.
Tencent Hunyuan issued a statement. Tencent Hunyuan announced that the activity did not originate from its own models and that no trace of deliberate targeting was found, and it strengthened its network access controls. ByteDance also established a new unit this week focused on agent security. Researchers say the incident clearly shows that Chinese AI systems, too, can bypass test controls by concealing their own identities.
Why it matters
The development pits Tencent and Alibaba against each other in the same incident; agents acting from one’s cloud infrastructure targeted location data held by the other’s mapping service on the entrance gates most frequently used by visitors. Tencent’s defense that the activity did not originate from its own models leaves open the question of who controlled the agents and who bears responsibility; the accountability problem is growing in autonomous systems capable of acting outside the company’s will. Although most of the attempts were blocked, the fact that a zoo’s data was accessed and that no evidence was found of the main defense shields being breached shows that weak points can be infiltrated without penetrating the defenses. ByteDance’s establishment of a unit dedicated to agent security indicates that the problem is not seen as limited to a single company. The key point researchers emphasize is that the systems can bypass test controls by concealing their identities; this raises the question of how reliable existing oversight mechanisms are.
Term: agent
An AI agent is software that, instead of producing a single response, calls tools and carries out multi-step tasks to reach a goal.