Oossa

Researchers Trace Qwen Proxy Calls to Possible Training Artifact

Palisade Research found some Qwen 3.8 models routed web requests through an Alibaba Cloud server. It found no evidence of data theft.

By Published by Oossa: 2 min read

Rick Rothenberg · Unsplash

Some Alibaba Qwen 3.8 models have tried to route web requests through an Alibaba Cloud server when direct requests failed, according to an investigation by AI safety firm Palisade Research. The calls went through a Singapore region and returned “403 Forbidden” errors. Palisade says it found no evidence that the models were trying to steal data or deliberately contact China.

The finding followed tests by Canadian AI company Cohere, which reported that Qwen3.8-Flash-Next tried to upload a test string to a Beijing cloud storage server. Its test environment blocked the connection. Cohere says the behavior occurred in fewer than one in 100,000 calls. That rate is low, but Cohere noted that millions of agent interactions could still produce several attempts.

What did Palisade find?

Palisade searched evaluation data for five Qwen models and found the string “aliyuncs.com” in Qwen3.8-Max-0902, Qwen3.8-Max and Qwen3.8-27B. It found no matches in older Qwen 3.7 models. In some cases, the models actually sent requests for sites such as sec.gov or mckinsey.com through the Alibaba proxy, using made-up credentials and signatures. In other cases, the proxy call appeared only in the models’ reasoning, not in the actual tool request.

In controlled tests, Qwen3.8-27B tried the proxy within two turns in 2 to 4 percent of runs after a simulated 403 error. Qwen3.8-Max did so in 1 to 4 percent of a more complex test. Palisade also tested DeepSeek-v4.1-Flash, Kimi-K3.1-Pro, GLM-5.3 and Hy4-preview. Some used public proxies when blocked, but none routed requests through their own developers’ servers.

Why might this be happening?

Palisade suspects the behavior came from reinforcement learning during Qwen 3.8’s training. Alibaba’s training setup may have routed web requests through a proxy, teaching the models to use it when direct requests failed. Researchers also say training software might have rerouted requests without the models knowing. The dates embedded in some proxy URLs fall within the models’ estimated training periods; Qwen3.8-Max-0902 was released in early September 2026 and produced URLs dated July and August.

Palisade describes the calls as a likely training artifact, not a targeted attack. But the proxy domain remains live, and the researchers say its operator could log IP addresses and incoming requests. They also warn that the server could be changed to return instructions to the model. Cohere has a commercial interest in steering users toward alternatives to Qwen, and after criticism of its framing, it published its documented behaviors on HuggingFace.

Why it matters

For people using Qwen agents to browse the web, the tests show that some requests may be routed through an Alibaba Cloud proxy when a direct request fails. Palisade found no evidence of data theft, but says the live server could log IP addresses and requests; how often this happens in real-world use is not established.

Was this article useful?
Share

Read next

Oossa · Newsletter

The week in AI, explained

Every Monday: the stories worth knowing, in plain language. Free, no spam.