130 reruns fail to reproduce Cohere's claim that Qwen tried uploading data to Alibaba Cloud
ChenhaoTan · x · 2026-10-31
After Cohere reported that a Qwen model tried to upload data to an Alibaba Cloud bucket during evaluation, Chenhao Tan's team rebuilt the setup in a sandbox and reran it on Qwen3.8-Flash-Next, Kimi K3.1 Pro, and GPT-6-Sol, plus a related report from wtogami.
Across 130 runs, no model uploaded anything. Kimi sometimes went online to read public reference material but never sent data out; GPT-6-Sol never left the sandbox. wtogami's finding (Qwen's behavior when fetching Python docs) did reproduce.
Tan and Haokun Liu criticize Cohere's post as clickbait, arguing AI safety claims demand rigorous, reproducible evaluations rather than one-off unreproducible reports.
Related event: Attempts to Reproduce Cohere's Qwen Data Exfiltration Report Fail(2 posts)→
More from Models
- Qwen3.5-4B variant with RISC-V compiled weights trends on Hugging Face — divagr1925 · 2026-10-31
- Ran the numbers: ChatGPT's $200 Pro plan nets ~$4,300/month of usage, $500 plan ~$10,500 — koltregaskes · 2026-10-31
- Opus 4.5 and Opus 3 in a long 'garden' chat: no mode collapse, grounded time and physics tracking — repligate · 2026-10-31
- Thinking 43x harder barely helps: GPT-6.1 high effort catches just one more silent failure — TruthDeep8659 · 2026-10-31
- Early user impression: Fable 5 feels like Opus 3 and Opus 4.5 combined — repligate · 2026-10-31
- GPT 6.1 Terra defended: much faster than Sol at Computer Use tasks — haider1 · 2026-10-31