Opening. We are looking for a research-oriented engineer to develop the methods that make our safety evaluations representative, robust, and informative. You'll work on questions like: How do we measure whether a model is safe? How do we create evaluations that reflect real-world usage rather than synthetic benchmarks? How do we know our graders are accurate?
What you'll do
This role sits at the intersection of applied ML research and engineering. You'll design experiments to improve how we evaluate model behavior, then ship those methods into pipelines that inform model training and deployment decisions. Your work will directly shape how Anthropic understands and improves the safety of our models across misuse, prompt injection, and user well-being.
What you need
- Design and run experiments to improve evaluation quality—developing methods to generate representative test data, simulate realistic user behavior, and validate grading accuracy
- Research how different factors (multi-turn conversations, tools, long context, user diversity) impact model safety behavior
- Analyze evaluation coverage to identify gaps and inform where we need better measurement
- Productionize successful research into evaluation pipelines that run during model training, launch and beyond.
- Collaborate with Policy and Enforcement to translate real-world harm patterns into measurable evaluations
- Build tooling that enables policy experts to create and iterate on evaluations
- Surface findings to research and training teams to drive upstream model improvements
