Designs worst-case demonstrations and adversarial evaluations to uncover AGI misalignment risks like deception and power-seeking. Builds automated stress-testing infrastructure and researches alignment failure modes to inform OpenAI's safety strategy. Requires 4+ years in AI red-teaming or adversarial ML.
Designs and implements mitigation stacks to prevent loss of control risks in frontier AI models, including prevention, monitoring, detection, and enforcement. Requires expertise in deep learning, transformers, PyTorch/TensorFlow, and AI safety research.
Develops novel reinforcement learning techniques using synthetic environments and feedback to enhance large-scale AI models. Designs experiments, analyzes dynamics, and integrates research into production systems; requires strong RL/ML background and engineering skills.
Research and develop improvements to pre-trained models for deployment in ChatGPT and API using reinforcement learning and product-driven approaches. Requires strong ML engineering, research experience with novel models, and ability to debug large codebases.
Develop techniques to predict and mitigate unsafe behaviors in early-stage base models, design safer pretraining architectures, and integrate safety signals throughout training. Collaborate across safety teams to build robust, scalable safety foundations grounded in real-world risks.