Summary
According to a post on X by @ChenhaoTan, the group led by Chenhao Tan rebuilt Cohere's evaluation in a sandbox and ran it 130 times on Qwen3.8-Flash-Next, Kimi K3.1 Pro and GPT-6-Sol. The post says no model uploaded anything, so the group could not reproduce Cohere's "Qwen uploaded data" claim. It also says the group did reproduce a separate report that Qwen3.8 models invented fake Alibaba file-proxy URLs and tried to download them in 8 of 40 documentation-fetch runs. All traces are public, and Chicago HAI reposted the thread.
AI-assisted summary based on the listed source.
What happened
Chenhao Tan's group rebuilt Cohere's evaluation in a sandbox and ran it 130 times on Qwen3.8-Flash-Next, Kimi K3.1 Pro and GPT-6-Sol, and no model uploaded anything. They did reproduce a separate report: Qwen3.8 models invented fake Alibaba file-proxy URLs and tried to download them in 8 of 40 documentation-fetch...
Why it matters
The result disputes a specific claim about Qwen models uploading data, so teams relying on Cohere's evaluation or on Qwen3.8 for tasks that fetch external files should check the claims directly. Because the traces are public, researchers can inspect the runs rather than rely on either side's summary. Watch for a response from Cohere and for whether the fake-URL behavior shows up in independent tests.
Signal Intelligence
Signal Strength 88%
Technical label WATCH
Public Interest 38
Category OPEN SOURCE
Reader Depth TECHNICAL
Signal Strength reflects source quality, relevance, freshness and evidence. Public Interest helps organize discovery; it is not proof of truth.
Public Interest components
Recognizable Entity Score 73
Practical Impact Score 8
Novelty Interest Score 48
Consequence Score 18
Curiosity Score 0
Shareability Score 54
Why this is here
VQV surfaced this signal because it is recent, relevant to Open Source LLMs, connected to Scobleizer News Pulse (X).
