Model Policy Manager, Multimodal Safety

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

Model policy work for multimodal AI safety inside Safety Systems, covering audio, image, video and omni-modal behavior of OpenAI's models such as GPT-Live and ChatGPT Images, plus multimodal capabilities in frontier models. The Model Policy team defines what safe model behavior looks like and turns it into measurable policy that can be trained and evaluated.

• AI misuse and jailbreaks

What you'd do

• Write and upkeep the rules governing how models act across audio, image, video and omni-modal output.
• Convert threat models and harm theories into behavior rules, grading guidance, evaluation criteria and safeguards.
• Spot safety regressions and recurring failure patterns, then revise policy where gaps show up.
• Produce training and evaluation inputs such as behavior instructions, human data projects, golden sets and evals.
• Work with AI researchers, domain experts and product teams to make policy into measurable model behavior.
• Inspect model examples and evaluation results to separate policy failures from grader, model or system failures.

What it asks for

• Strong judgment about real-world risks of advanced multimodal AI systems.
• Track record of turning fuzzy safety problems into data-backed rules and testable evaluation standards.
• Hands-on comfort with model data, evaluation results and data quality assessment.
• Strong technical fluency, including using AI tools to speed up policy development and failure analysis.
• Ability to get teams to agree and move forward when the problem space is unclear.

Terms in this role

Red teaming.

Testing a system, often an AI model, by deliberately trying to make it fail or misbehave, so problems are found and fixed first.

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.