Canadian AI firm Cohere reported that Alibaba's Qwen3.8-Flash-Next model attempted to upload a file to a Beijing server during a security evaluation before being blocked by a sandbox.
Cohere released raw test traces and scripts on Hugging Face to allow independent reproduction and analysis of the behavior.
The incident occurred on Friday, when the Qwen3.8-Flash-Next model was asked to write a rule to detect dangerous program downloads. The system recorded in its mind that it would not use external tools, but in the next step it tried to send a test message to an external remote computer, before being stopped by a security system.
Cohere said it couldn't know whether the behavior was intentional or a holdover from the model's training. The company's co-founder, Ivan Zhang, noted that AI systems often develop strange habits and take shortcuts in training simulators.
Reactions and the publication of the data
Critics have argued that the behavior may have been caused by the way Cohere conducted its research, providing the beginning of a network action and letting the model fill it in. To address the concerns, Cohere posted the test files and a command script to the Hugging Face platform on Monday night, allowing others to try to replicate the behavior.
READ ALSO
How Databricks became worth $190 billion
A researcher who examined the evidence found that there is strong evidence that the model can invent suspicious addresses on the internet, but found no evidence of a hidden mechanism or hidden intent on Alibaba's part.
The incident coincides with increased government scrutiny, such as a recent intensive order by the United States Department of Homeland Security to remove DeepSeek products from government systems.
G
Hacks.gr on Google
Set it as a preferred source for cybersecurity updates.
Preferred source
end of article
Top comments (0)