The Safety Training research team aims to fundamentally advance our capabilities for precisely implementing safe behavior in AI models.
Key focus areas include how to train nuanced safety behaviors, how to make the model robust to bad actors, how to address privacy and security risks, and how to make the model trustworthy in safety-critical situations.
About the role
We're seeking a researcher to train and evaluate models for U.S. government use, with a focus on national security applications.
You'll advance safety post-training and robustness, helping models follow nuanced policies while preserving their usefulness and capabilities.
Responsibilities
Research and implement methods for safety training, reinforcement learning, and adversarial robustness.
Develop evaluations, identify model failure modes, and use findings to improve training.
Work with research, engineering, security, and policy partners to support safe, reliable deployment.
Requirements
Bring 4+ years of relevant AI safety research experience, including RLHF, adversarial training, or robustness.
Have a degree in computer science, machine learning, or a related field, and strong deep learning research or engineering skills.
Have experience improving model safety for deployment and enjoy collaborative research.
Are motivated by OpenAI's mission and the responsible use of AI in safety-critical settings.