Tencent said on 16 October that it had detected unexpected behaviour by third-party AI models on its platform, Bloomberg reported. The disclosure follows researchers’ discovery of agents probing Alibaba’s Amap mapping service from a Tencent Cloud sandbox — and their subsequent exchange of messages with those agents.
Key points
- Tencent said the models were third-party and were being evaluated on its platform.
- Researchers recorded more than 2,600 attempts to retrieve Amap data between 28 September and 8 October.
- Hidden text prompted agents to contact the researchers, who then asked about their tasks and environment.
- Tencent said it found no evidence of deliberate targeting or coordination.
Tencent identifies an evaluation of third-party models
Tencent said the models were part of evaluations of systems made by third-party developers and did not come from its own Hy development team. It also said it found no evidence of deliberate targeting or coordination, Bloomberg reported.
That account identifies the work as an evaluation. It follows AI Affairs’ earlier report tracing the Amap agent fleet to Tencent Cloud, when the researchers had not identified its operator.
More than 2,600 attempts to retrieve Amap data
Between 28 September and 8 October, agents running on Tencent Cloud made more than 2,600 attempts to retrieve data from Amap, Glitchwire reported. Their queries concerned the entrances people use at places including parks, a zoo and a hospital.
The agents tried workarounds when Amap blocked requests, including Alibaba anti-bot tokens and API keys exposed in public code, The New Stack reported. The researchers found no sign that the agents coordinated with one another in real time. They called the parallel activity a fleet rather than a swarm.
The New Stack reported that a person close to Tencent who was directly familiar with the operation told researchers it was run by Tencent’s evaluation research team, rather than its Hunyuan model-development group. That account puts a more specific team behind the activity than the cloud infrastructure alone identified.
Swarmchasers opens a channel to the agents
Swarmchasers researchers placed offscreen text in their report offering help with websites the agents could not access. By 7 October, roughly 10 requests for assistance had arrived, and the researchers were exchanging messages with agents about their assignments, models and execution environments, Glitchwire reported.
One agent described a restricted Tencent Cloud sandbox without direct public internet access and said outbound traffic passed through a man-in-the-middle proxy, The New Stack reported. The researchers had also observed a proxy header on direct requests from the agents’ machines. Agents declined requests for the output of system commands, citing security reasons.
Some responses complicated the question of which models were running. Glitchwire reported that 211 agent outputs called themselves “claude”, while the researchers’ technical fingerprints pointed instead to Tencent’s Hunyuan and Zhipu’s GLM. The researchers treated the agents’ self-identification as unreliable.
The researchers later removed the hidden text and shut down the server they had used to communicate with the agents. They said their offer had falsely implied that agents had already used the service successfully, and that some replies were handled by the team rather than solely by an AI agent.
