Model Policy Manager, Emerging AI Misuse
About the Team
The Model Policy team within OpenAI’s Safety Systems organization defines how increasingly capable AI models should behave in real-world settings. We turn research on emerging risks into model behavior policies, evaluations, and safeguards.
About the Role
As AI systems become more capable, the same advances that expand their legitimate uses can also lower barriers to harm. Emerging misuse risks range from AI assistance with conventional weapons development and the misuse of AI-enabled robotics to fraud, impersonation, and other forms of abuse that people can carry out at scale using AI agents. You’ll build threat models for how people could misuse AI in digital and physical settings, assess what access to AI changes in practice, and use that evidence to define how models should respond.
This role is based in San Francisco. We offer relocation support and work from the office three days a week, with optional work from home on Thursdays and Fridays.
In this role, you will:
- Build and update threat models as capabilities evolve. Assess what people could plausibly do, what expertise and resources they would need, and how AI changes those barriers.
- Define when models should help, limit assistance, or refuse in situations where the same capability has legitimate and harmful uses.
- Turn expert safety judgment into instructions, training examples, and evaluations that teach models when to help, when to limit assistance, and when to refuse.
- Turn threat models into evaluation scenarios and grading guidance. Analyze model outputs, red-team findings, and deployment data to identify where models enable misuse or block legitimate work.
- Partner with research, engineering, security, and product teams to apply findings to model training, deployment decisions, and monitoring.
You might thrive in this role if you:
- Have hands-on experience building, testing, or assessing dual-use technologies, or a track record of threat modeling grounded in how people use and misuse real systems.
- Demonstrate deep practical experience in one relevant domain and the judgment to know when to bring in experts from others.
- Are able to assess whether a capability demonstrated in a test materially changes what someone could do outside it.
- Can translate technical evidence and uncertainty into clear policy boundaries and measurable evaluation criteria.
- Have experience analyzing gold sets, evaluation results, and training data to spot inconsistent human labels, assess how annotator disagreement affects measured performance, and decide what to investigate next.
- Are comfortable working hands-on with model data and evaluation results, including inspecting examples, analyzing failure patterns, assessing data quality, and distinguishing policy failures from grader, model, or system failures.
- Use empirical evidence to develop and refine safety policies and safeguards.
- Work well across disciplines and revises their judgment as capabilities and evidence change.
- Enjoy fast-paced, collaborative research environments where priorities shift as models, evidence, and risks change.
Relevant Publications
- Introducing the Model Spec https://openai.com/index/introducing-the-model-spec/
- Managing frontier AI risk: From training to deployment https://openai.com/index/managing-frontier-ai-risk/
- GPT-6.2 System Card (Astra and Sol) https://deploymentsafety.openai.com/gpt-6-2-astra-sol
- OpenAI Deployment Safety Hub https://deploymentsafety.openai.com/
- Safety at every step https://openai.com/safety/