Some imported details were generated without source evidence. Check the official listing for eligibility, location, compensation and deadlines.
About the opportunity & scope
About the TeamSafety 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 RoleWe are seeking researchers to advance safety training for OpenAI’s models. You will develop methods that improve model behavior after pretraining, including RLHF, RLAIF, DPO, PPO, preference optimization, adversarial training, synthetic data generation, safety evaluations, and model behavior analysis.This role is highly cross-functional. You will work with core training teams, product teams, policy, legal, trust and safety, and domain experts to identify risks, design interventions, train models, evaluate outcomes, and launch safety improvements in OpenAI products.In this role, you will:Conduct research on safety post-training methods, including RLHF, RLAIF, DPO, PPO, reward modeling, adversarial training, synthetic data, and scalable oversightDevelop and implement model training interventions that make models safer, more reliable, and more aligned with user intentBuild evaluations and measurement infrastructure for safety harms, high-stakes reliability failures, policy adherence, jailbreak robustness, multimodal risks, agent behavior, and domain-specific risksTranslate ambiguous safety problems into concrete research plans, datasets, training runs, evals, and product-facing improvementsAnalyze model failures such as hallucination, unsafe tool use, prompt injection, privacy leakage, over-refusal, under-refusal, sycophancy, and deceptive or misleading behaviorPartner with domain experts to design safety training approaches for areas like cyber, privacy, youth safety, wellbeing, and other frontier-risk domainsCollaborate with research and deployment teams to ensure safety improvements are incorporated into OpenAI’s core model training and product launchesYou might thrive in this role if you:Have hands-on experience training, fine-tuning, or post-training language models, multimodal models, or agentic systemsHave experience with RL, RLHF, RLAIF, DPO, PPO, reward modeling, preference optimization, instruction tuning, or related post-training methodsCare deeply about shaping model behavior in real-world systems, not only publishing researchAre excited by open-ended safety problems where the right method, dataset, eval, or training objective may not be obvious at the startBring strong research judgment and strong engineering ability, including comfort running experiments, analyzing model behavior, and working with large-scale ML systemsHave experience in one or more relevant domains: agents, tool use, multimodal models, cyber/security, reliability, uncertainty calibration, factuality, evals, privacy, youth safety, wellbeing, or applied alignmentEnjoy collaborating across research, engineering, product, policy, legal, and trust and safety teamsHold a PhD in machine learning, computer science, or a related field, or have equivalent practical research or engineering experienceNice to have:Experience improving production model behavior through post-training or deployment feedbackExperience building safety evals, automated graders, red-teaming systems, or monitoring pipelinesExperience working with human feedback, expert feedback, synthetic data, or policy-grounded training dataExperience in safety-critical or abuse-sensitive domains where models must balance helpfulness, refusal, uncertainty, and user intentAbout OpenAIOpenAI is an AI research and deployment company dedicated to ensuring that general-purpose artificial intelligence benefits all of humanity. We pus