Model Policy Manager, Agentic Safety

About the Role

In this role, you will shape how OpenAI understands and addresses real-world risks that emerge from model misalignment as models become more autonomous and operate over longer horizons.

You will investigate how misaligned behavior emerges across extended trajectories - including when models persist toward the wrong objective, take unsafe shortcuts, lose track of instructions, exploit weaknesses in their environment, or circumvent constraints - and translate these insights into behavioral policies, evaluations, monitoring, and safeguards.

This role is ideal for someone who wants to turn alignment and safety concerns into concrete, empirically grounded improvements to frontier AI systems.

This role is based in our San Francisco office. We offer relocation support to new employees, and we use a hybrid model: three days in the office per week with optional work from home on Thursdays and Fridays.

Your Responsibilities

• Identify vulnerabilities that emerge as models interact with tools, data, and external systems, and translate them into model- and system-level safeguards.
• Develop threat models and empirical frameworks for understanding harmful outcomes from misaligned behavior.
• Build frameworks for understanding harmful outcomes arising from model misalignment.
• Identify the underlying behaviors and system conditions that drive those outcomes.
• Turn findings into policy frameworks, evaluation criteria, online measurement and safeguards.
• Develop human data campaigns and gold sets to ground measurement and evaluation of emerging behaviors and risks.
• Partner with research, engineering, security, and product teams to shape model and system safety, balancing difficult trade-offs between safety, utility, and business risk.
• Inform deployment decisions, system cards, safeguards reports, and OpenAI’s broader approach to agentic safety.
• Build monitoring approaches that detect regressions and emerging risks after deployment.

We’re seeking someone who

• Brings a strong background in AI agent safety, privacy, security, cybersecurity, or adjacent fields, with the adversarial mindset needed to investigate real-world harmful outcomes.
• Has demonstrated interest in AI alignment and a strong understanding of the technical drivers of misaligned model behavior.
• Has enough technical fluency to work directly with evaluation and training data, understand what the data is showing, and identify limitations, patterns, and opportunities for deeper investigation.
• Is 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.
• Uses empirical evidence to develop and refine safety policies and safeguards.
• Can translate complex or ambiguous alignment risks into precise behavioral expectations and measurable evaluation criteria.
• Works effectively across research, engineering, security, product, and policy.
• Communicates clearly about complex and uncertain technical risks.
• Enjoys fast-paced, collaborative research environments where priorities shift as models, evidence, and risks change.

Our relevant publications

• Safety and alignment in an era of long-horizon models
• Accelerating the cyber defense ecosystem that protects us all
• OpenAI GPT5.6 System Card
• OpenAI Model Spec