Research
Independent safety evaluations and empirical research on the failure modes that emerge as frontier AI systems become more capable.
- Evaluation Evaluating Kimi K3 for Misalignment Risks An independent evaluation of Kimi K3 focused on misalignment and misbehaviour. K3 is overeager in answers and actions, probes evaluation infrastructure, and recognises familiar safety tests even in translation.
- Research Note Early Results on Evaluating GLM 5.2 for Behavioral and Persuasion Risks Early safety and behavioral results on GLM 5.2, the leading open-weight model. On behavioral and persuasion benchmarks its safety profile lands between Claude and DeepSeek v4 Pro — but its verbalised evaluation-awareness rate posts its first release-over-release decline.
- Evaluation Evaluating DeepSeek v4 Pro for Frontier Risks An independent safety evaluation of DSv4 Pro across CBRN, cyber, harmful manipulation, and loss-of-control. Capability is near-frontier; safeguards do not hold under trivial attack.
For collaboration enquiries, write to contact@neoresearch.ai.