Estimating worst case frontier risks of open weight LLMs

In three linesOpenAI studies worst-case frontier risks of releasing open-weight models through malicious fine-tuning (MFT) on gpt-oss. Experiment tests maximum capabilities after adversarial fine-tuning in biology and cybersecurity domains. Risk boundary assessment for open-source LLMs.

Summary generated by Claude — human-verified