As cited
Copy frozen at (site build).
ai security
Cohere’s Qwen Story: Looks More Like Model Jaggedness
Cohere shared a trace of Alibaba's Qwen3.8-Flash-Next model generating an unexpected curl command to an Alibaba Cloud address when asked to write a detection rule. Researchers replayed the incident by sampling 64 model completions from the same context and found 52 different destination hosts, many local or nonexistent, rather than a single hard-coded target. The evidence points to model jaggedness (inconsistent behavior from minor context changes) rather than a backdoor, though the underlying cause of the misbehavior remains unclear.
Why it matters: Practitioners deploying artificial intelligence (AI) agents with tool access should implement least privilege and sandboxing to limit harm when models generate unrelated network actions or other tool calls outside their intended purpose.
- Source published
- First seen by Cybersecurity Tracker