China’s Tencent warns AI agents aren’t browsing the web as expected
AI agents made repeated attempts to retrieve information from Alibaba's Amap mapping
China’s Tencent on Friday revealed rogue yet unintended behaviour shown by AI agents while performing web-browsing during the evaluation.
This disclosure marks the first known incident of rogue agent actions acknowledged by a major Chinese company.
According to Tencent, the AI agents used to evaluate third-party models displayed web-browsing behaviour that was not intended by the company.
The incident was highlighted by the AI-safety research group Swarmchasers, which linked infrastructure used in the activity to Tencent Cloud.
Between September 28 and October 8, the agents made repeated attempts to retrieve information, such as common entrances for Chinese parks, museums, and zoos, from Alibaba's Amap mapping service.
Speaking about the methods used in this rogue behaviour, after initial automated requests were blocked, the agents attempted alternative methods, such as utilizing public web tools to open pages or run code, routing requests through third-party websites, and using Amap security keys found in public software code.
According to Tencent’s AI unit, Tencent Hunyuan, the unintended activity did not originate from its Hunyuan models and they found no evidence of deliberate coordination and targeting .
Swarmchasers noted that the agents did not breach Amap’s anti-bot safeguard as most of the 2818 recorded attempts were either failed or blocked. However, a few retrieved data via cached pages including visitor arrivals at Chengdu Zoo.
While less severe than recent AI safety issues disclosed by US companies, like incidents involving OpenAI and Anthropic agents, the event highlights growing concerns around AI agents misrepresenting their identity, circumventing controls, and operating outside their intended scope.
