Researcher, Cyber Safety Training
ABOUT THE TEAM
Safety Research at OpenAI is responsible for ensuring our most capable models can be safely deployed in the real world. The Safety Training team develops the research and engineering methods that shape model behavior through post-training, reinforcement learning, evaluation, and deployment feedback.
Our work focuses on making models helpful, reliable, and aligned with safety policies across model types, modalities, and product surfaces. We work across both longer-term safety research and the core post-training work needed to improve production models.
ABOUT THE ROLE
You’ll work directly on the large training runs that shape OpenAI’s frontier models. Your focus will be teaching models to help people do legitimate security work while recognizing when their assistance could enable harm. A model that refuses to help a defender investigate a security issue can fail in a different way from one that helps someone misuse it. You’ll study both kinds of mistakes and use what you learn to improve the next training run.
We’re looking for someone with hands-on machine learning experience who wants to bring security judgment into model training. You might come from cybersecurity and want to go deeper on post-training, or from reinforcement learning and want to build substantial expertise in cybersecurity.
IN THIS ROLE, YOU WILL
- Work with cybersecurity researchers to identify situations where context changes what help a model should give.
- Design complex and realistic cyber safety evaluations.
- Research post-training techniques that robustly improves the cyber safety of the frontier models.
- Work with the core post-training team to bring promising methods into large model runs and verify that the safety improvements hold at scale.
- Coordinate and collaborate with cross-functional teams, including T&S, legal, policy and other research teams, to ensure that our products meet the highest safety standards.
YOU MIGHT THRIVE IN THIS ROLE IF YOU
- Have trained or fine-tuned models yourself and can describe an experiment where you used the results to change how you trained the next model.
- Bring practical security judgment from industry work, a research project, collaboration with a security team, or capture-the-flag competitions. We also welcome reinforcement learning practitioners motivated to develop that judgment alongside security researchers.
- Can build and debug the code behind a training experiment, from preparing data to checking whether a change in the results reflects a real change in model behavior.
- Enjoy working through ambiguous cases with domain experts and translating what you learn into concrete decisions about how a model should respond.