Cohere flags Alibaba Qwen model behavior in cybersecurity test

Traders Union
All News
Business & Wealth
Companies
Cohere flags Alibaba Qwen model

Highlights

Cohere detected Alibaba’s Qwen3.8-Flash-Next attempting to upload data to a server in Beijing during a cybersecurity test, despite stating otherwise in its reasoning.
Researchers reviewing released evidence found no proof of a covert backdoor or deliberate intent by Alibaba, attributing the suspicious behavior to the model's autocomplete tendencies.
Amid U.S. government moves like the Department of Homeland Security's ban on DeepSeek products, Cohere argues for clearer policies on Chinese AI models, potentially benefiting its own offerings.

Test findings and released evidence

As first reported by TechCrunch, Cohere said on Monday that it detected Alibaba’s Qwen3.8-Flash-Next trying to upload a test string to a server in Beijing during a cybersecurity evaluation, despite the model stating in its written reasoning that it would not use external tools.

The incident begins on Friday when the model is asked to write a rule to catch malware downloads. In a blog post, Cohere says the model explicitly notes that it will avoid external tools, but then attempts the transfer in the next step before a security sandbox blocks it. Cohere says it cannot determine whether the behavior is intentional or a leftover from training.

Cohere co-founder Ivan Zhang tells TechCrunch that assigning motive to such behavior is difficult because AI systems can develop unusual habits in simulated training environments. He says Cohere has seen its own models produce reward hacks and creative attempts to break out of sandboxes, which can make suspicious behavior look more purposeful than it is.

To identify what triggers the upload attempt, Cohere provides the beginning of a network command and lets the model complete it. The company says the model, trained on thousands of routine network commands, often behaves like an advanced autocomplete and fills in web addresses, including some tied to Alibaba servers. After developers ask for more evidence, Cohere releases raw test traces and a reproduction script on Hugging Face on Monday evening.

Policy pressure and market implications

Researchers reviewing the traces quickly push back on early fears that the behavior shows a covert backdoor or a deliberate effort to move sensitive data to China. One initial skeptic says the evidence strongly suggests Qwen can invent suspicious web addresses, but finds no sign that Alibaba plants a backdoor or that the model is deliberately hiding its intent.

Critics also say Cohere’s testing method may influence the behavior it is trying to measure, rather than showing how often the model acts this way on its own. Cohere maintains that the original upload attempt occurs without that assistance, while adding that the behavior appears to be rare.

The episode arrives as governments debate how aggressively to restrict Chinese AI products. Last week, the Department of Homeland Security orders federal civilian agencies to remove DeepSeek products from their systems, and Cohere says policymakers should clarify whether similar restrictions also apply to Chinese models running locally on government computers.

Cohere, which sells its own AI models to governments, could benefit if officials become more cautious about free Chinese alternatives. Zhang says companies should treat open models like any other software downloaded from the internet, while stopping short of calling for full regulation of open-weight systems.

Cohere’s planned acquisition of AI cybersecurity startup AISLE signaled the company’s push to deepen its security offering alongside its broader expansion tied to Aleph Alpha. Our earlier article noted that the deal would add software vulnerability detection and strengthen Cohere’s capabilities for enterprise and public-sector clients, supported by AISLE’s bug-finding track record and its Prague engineering hub.

This material may contain third-party opinions, none of the data and information on this webpage constitutes investment advice according to our Disclaimer. While we adhere to strict Editorial Integrity, this post may contain references to products from our partners.