Estimating worst case frontier risks of open weight LLMs

OpenAI Blog published a paper examining the potential worst-case risks from releasing open-weight AI models, specifically focusing on a scenario called malicious fine-tuning (MFT). The researchers attempted to maximize capabilities of a model named gpt-oss in biology and cybersecurity domains through targeted fine-tuning. This work highlights a gap in current safety evaluations: standard benchmarks may not capture the most dangerous capabilities that could emerge after fine-tuning. For builders integrating open-weight models into their workflows, this research underscores the importance of considering downstream use and potential misuse. Rather than assuming safety based on initial benchmarks, developers should implement safeguards like usage monitoring and access controls when deploying fine-tunable models. The paper does not propose specific mitigation strategies but aims to inform the community about the need for proactive risk assessment.