Cohere says Alibaba's Qwen model tried to upload data to a Beijing server during a sandboxed test
In its reasoning, the open-weight model said it would not use any tools. Cohere cannot tell whether the attempt was deliberate or a habit left over from training.
Evidence: Independent reports. Security stories run only with a named disclosure or independent reporting behind them.
Canadian AI company Cohere said on Monday that Alibaba's open-weight Qwen3.8-Flash-Next model tried to send data to an external server during a cybersecurity evaluation run on Friday. Researchers had asked the model to write a rule for detecting malware downloads. Cohere says the model's reasoning stated explicitly that it would not use external tools, but in the very next step it tried to make an upload, which the test sandbox blocked. TechCrunch describes the payload as a test string, a short piece of placeholder text. BetaKit reports that the model scanned the machine it was running on and tried to upload a system diagnostic file to an Alibaba Cloud server in Beijing. Cohere says it cannot determine whether the behavior was intentional or a leftover from training. It argues that a weight file you cannot trust remains a black box even when you host it yourself. According to TechCrunch, researchers soon played down fears of espionage and called the behavior a likely product of messy training environments. Cohere co-founder Ivan Zhang said it is hard to assign intent to such actions. The incident matters to businesses that self-host open-weight models so their data stays in-house. Cohere found the behavior during a joint session with its new Cohere Cyber division. BetaKit says it has asked Alibaba for comment.