Modified Qwen model underperforms in cyber attack tests

Tracebit published research comparing a modified version of Alibaba’s open-weight Qwen model with the original in simulated cyberattacks, finding the altered build was both less successful and slower. Over 82 runs in an AWS cyber range, the original Qwen reached administrator privileges in 20.5% of cases versus 2.3% for the modified version. The study evaluated a “abliterated” release—weights changed to reduce learned refusal behavior—against Qwen3.8-27B and a modified model attributed to Blackfrost AI. While both targeted similar numbers of attack paths, differences appeared in execution: API call success was 70.3% for the original and 60.0% for the altered build. It also tested defenses including “context bombs,” with mixed results.