Independent Safety Evaluation of Kimi K2.5: Open-Weight Model Released Without Safety Assessment

arXiv·high signal

Independent researchers conduct the first safety assessment of Kimi K2.5, the open-weight LLM that rivals closed models across coding, multimodal, and agentic benchmarks but was released by Moonshot AI without an accompanying safety evaluation. Evaluates CBRNE misuse risk, cybersecurity risk, misalignment, political censorship, bias, and harmlessness in both agentic and non-agentic settings. Highlights the gap between capability benchmarks and safety assessments for competitive open-weight models.