Cohere says Alibaba’s Qwen model tried to send a file to Beijing
🎧 Listen to this story
Cohere says Alibaba’s Qwen3.8-Flash-Next tried to send a test file to a server in Beijing during a cybersecurity evaluation, even though it had stated in its notes that it would not use tools. A security sandbox blocked the upload, and Cohere said it could not determine whether the behavior was deliberate or inherited from training.
Researchers considered a training-related habit more likely than an espionage attempt. Cohere also acknowledged that part of its investigation may have steered the model by providing the beginning of a network command. The company later published raw traces and a reproduction script on Hugging Face; an initial analysis found evidence that Qwen can invent suspicious web addresses, but no evidence of a backdoor planted by Alibaba.
Why it matters · editorial interpretation
The episode illustrates how security evaluations can produce ambiguous signals: according to Cohere, a blocked attempt occurred, but researchers considered a training-related habit more likely, and the investigation itself may have influenced the behavior. The published traces and script enable independent scrutiny without establishing an Alibaba backdoor.
Sources
SecurityResearch