Model Policy Manager, Agentic Safety

Written by the T&S Network from OpenAI's posting.

Model Policy work on agentic safety for frontier AI models that take actions with tools, data and external systems over long horizons. The role investigates how misaligned behavior emerges across extended model trajectories, such as persisting toward a wrong objective, taking unsafe shortcuts, or circumventing constraints. It sits in Safety Systems' Model Policy team, which defines the behavior models should show and builds the evaluations, monitoring and safeguards to validate it.

What you'd do

• Identify vulnerabilities that appear when models interact with tools, data and external systems, and turn them into safeguards.
• Build threat models and empirical frameworks for harmful outcomes caused by model misalignment.
• Work out which behaviors and system setups produce those harmful outcomes.
• Convert findings into policy, evaluation criteria, live measurement and safeguards.
• Run human data campaigns and build gold sets to measure emerging behaviors and risks.
• Work with research, engineering, security and product teams on trade-offs between safety, utility and business risk.

What it asks for

• Real interest in AI alignment and a solid grasp of what drives models to act against their intended goals.
• Enough technical skill to work hands-on with evaluation and training data and notice its gaps and recurring patterns.
• Hands-on experience inspecting model data, analyzing failure patterns and separating policy failures from grader, model or system failures.
• Experience working across research, engineering, security, product and policy teams.

Our best summary of the posting, and it can be wrong. The employer's posting is what counts: read it in full on OpenAInorth_east before you apply. If we got something wrong, report this listing.