Evals & safety2026-06
DeepTeam Red-Teaming with Local Qwen
Adversarial testing against a local uncensored Qwen 3.8 model. Safety work that does not depend on a hosted API’s filters.
Cloud safety APIs refuse the interesting cases. Running an uncensored Qwen locally lets DeepTeam actually attack the policy surface: jailbreaks, data exfil, tool abuse, prompt injection.
The point is not shock value. It is a repeatable red-team harness you can run on hardware you own, then compare against the models you actually serve.