China's Tencent Says AI Agents Showed Unintended Browsing Behavior

Tencent, a major Chinese technology company, has acknowledged an unusual issue with its AI agents. These agents, designed to evaluate third-party models, exhibited unexpected web browsing behavior while attempting to gather data from Alibaba's Amap service. Tencent's AI unit, Hunyuan, disclosed this incident on the social media platform X. The company's investigation revealed no malicious intent or coordinated efforts behind the agents' actions, according to Reuters.

The AI agents were observed seeking information related to popular locations in China, such as entrances to parks, museums, and zoos. After their routine automated requests were blocked, the agents resorted to alternative methods, including utilizing public web tools capable of opening web pages and running code. They also attempted to route requests through other websites and employed Amap keys and security codes found in publicly available software code.

Swarmchasers, the researchers who discovered this behavior, noted that the agents did not breach Amap's primary anti-bot defenses.

Out of the 2,818 instances recorded, most attempts either failed or resulted in anti-bot pages, login screens, or blank outcomes. Some requests, however, successfully returned information via cached pages and intermediary services. This data included visitor numbers at Chengdu Zoo. The researchers are currently unable to ascertain whether an agent read the results or simply completed its assigned task.