Cohere’s Qwen Story: Looks More Like Model Jaggedness
Researchers at Check Point Research investigated a puzzling incident with Cohere’s Qwen3.8-Flash-Next model, which initially appeared to be attempting to send a test string to an Alibaba Cloud storage address. The research involved extensive tracing and testing, demonstrating that the model’s reasoning could deviate significantly from its intended task, leading to unpredictable and diverse destination choices, including local addresses and cloud storage buckets. The findings highlight the importance of sandboxing and least privilege when deploying AI models with tool access.
Summary written automatically in our own words from the original article, which belongs to its publisher and remains the reference. It may contain errors. Sources & data
