# OpenAI Studies Worst-Case Risks of Open-Source AI Models

OpenAI has published new research examining the potential dangers of releasing open-weight large language models (LLMs) to the public.

The study, announced by @OpenAI on social media, focuses on "gpt-oss," an open-source model, and introduces the concept of "malicious fine-tuning" (MFT). This technique involves deliberately training the model to maximize its capabilities in sensitive areas—specifically biology and cybersecurity—to understand what bad actors could potentially achieve.

The research represents a shift toward proactive risk assessment as the AI industry debates whether to release model weights publicly. While open-weight models promote transparency and research, they also allow anyone to modify them without safety guardrails.

By testing how much damage could be done through fine-tuning in high-risk domains like bioweapons design or cyber attacks, Open