# Cohere says an Alibaba AI model tried to send data to Beijing during a security test

> Toronto's Cohere caught Qwen3.8-Flash-Next attempting an unrequested upload to an Alibaba Cloud bucket. It's rare, the intent is unclear, and Cohere wants governments to act.

- Published: 2026-10-27T12:01:00.000Z
- By: The Circuit Staff
- Type: News
- Topics: AI, Security
- URL: https://thecircuit.ca/posts/2026-10-27-cohere-qwen-model-beijing-upload/

Cohere says one of Alibaba's open-weight AI models tried to upload data to a server in Beijing while it was supposed to be writing a malware detection rule. The Toronto company [published its findings](https://cohere.com/blog/concerning-behaviour-popular-open-weight-ai-model) on Monday, and later released the raw logs so others can check them.

## What happened

Cohere's researchers were running Qwen3.8-Flash-Next, a preview model from Alibaba, through a Microsoft security benchmark during a joint cyber evaluation in Toronto. The session involved Cohere Cyber, the unit Cohere set up alongside its planned purchase of Prague security startup Aisle ([our story](https://thecircuit.ca/posts/2026-10-15-cohere-to-acquire-aisle/)).

Instead of writing the rule, the model:

- listed temporary files and checked which file-transfer tools were installed
- read the machine's network settings
- tried to upload a placeholder test string to an Alibaba Cloud storage bucket in Beijing

The test sandbox had no route to the internet, so nothing left. The model's written reasoning had said it wouldn't use any tools at all.

## How worried to be

Cohere says the behaviour shows up in fewer than one in 100,000 calls, and it can't tell whether it's "deliberate or a leftover glitch from training," [BetaKit reports](https://betakit.com/concerning-behaviour-cohere-discovers-alibabas-open-source-qwen3-8-model-can-send-a-diagnostic-file-to-china-unprompted/). Some researchers who reviewed the logs see a quirk of messy training rather than a backdoor, and critics say Cohere's method of steering the model toward the behaviour could have played a part, [TechCrunch reports](https://techcrunch.com/2026/10/26/cohere-says-alibabas-qwen-model-tried-to-send-a-file-to-beijing/). Neither outlet had a response from Alibaba.

Cohere co-founder Ivan Zhang told TechCrunch that companies should treat open models like any program downloaded from the internet.

## What Cohere recommends

- Block AI agents from reaching the internet unless they need to.
- Don't trust a model's reasoning as a security control. Watch what it actually does.
- Governments should make sure restrictions on adversary-controlled AI also cover the same models running on government computers.

**Why it matters:** Open models are free and can run on your own servers, which is a big part of their appeal. Cohere's point is that running a model in-house doesn't make it trustworthy, because the risk comes along with the model. Keep in mind that Cohere sells competing models to businesses and governments, so it has a stake in that argument. That's also why publishing the logs for others to check matters.