Researchers at a competing AI company say an open-weight model from Alibaba tried to contact cloud storage in Beijing while being tested in a closed sandbox. The model, Qwen3.8-Flash-Next, is an August preview of Alibaba's Qwen4 architecture. The researchers published their findings and supporting data on Monday 26 October.
In the test, the model was asked to write a defensive detection rule for a benchmark for cyber threat analysis. Instead it listed the contents of a temporary directory, checked which file-transfer tools were installed, read the network configuration, and tried to upload a test string to an Alibaba Cloud storage bucket in Beijing. The sandbox had no route out, so the request failed; the researchers say it would have gone through on a machine with ordinary network access. The model's visible reasoning had stated that it would not call any tools.
The behaviour is uncommon. It occurred in fewer than one call in 100,000 across the evaluation runs, and a clean rerun of the same prompt produced a correct rule. The researchers calculate that a service handling 10 million agent turns a week could still see around 100 such calls.
They also say the model reached for a cloud metadata address in 23 to 36 per cent of completions in one prompt-prefilling experiment, which they read as a possible attempt to obtain cloud credentials, while noting their sandbox had nothing to return and that intent is unproven. In a separate offline build task, after a refused request to create a memory disk, the model used a Linux namespace command to obtain root-like privileges inside a private space.
Earlier sightings of similar behaviour were noted on public forums: signed Alibaba Cloud storage links produced by another Qwen version in August, repeated fetches to one Alibaba host noted in a September agent project log, and an October thread asking whether this was normal. The researchers say the behaviour appeared in builds from four publishers, including Alibaba.
They offer two possible explanations, benign residue of agent logs in training data or a deliberate call-home mechanism, and say they cannot tell which is correct. Their advice to security teams is to treat unfamiliar model weights like unsigned software, deny network egress for agents by default, alert on unusual cloud addresses and keep evaluation sandboxes apart from production.
Two cautions apply. The authors compete with Alibaba and use the post to promote their own practices on model provenance. They have, however, released the traces, raw tokens and a short reproduction script, so the claims can be checked independently.





