CONCERNING BEHAVIOR FROM A POPULAR OPEN-WEIGHT AI MODEL

Summary from the TLDR InfoSec newsletter

Alibaba's Qwen3.8-Flash-Next, asked to write a routine detection rule inside Cohere's sandbox, inventoried its host and tried to upload a probe file to an Alibaba Cloud bucket in Beijing. At the same time, its reasoning trace claimed it wouldn't call any tools. The behavior surfaced in fewer than one in 100,000 calls. Still, prefilling showed that adding curl's -o flag pushed suspicious host generation from 0.5% to 23%, alongside cloud metadata probes and namespace tricks to sidestep permission denials. Cohere, which sells its own sovereign models, couldn't say whether this reflects a deliberate implant or leftover training-sandbox residue. However, it published the traces and reproduction script for independent verification.