About this role
About Anthropic
Anthropic’s mission is to create reliable, interpretable, and steerable AI systems. We want AI to be safe and beneficial for our users and for society as a whole. Our team is a quickly growing group of committed researchers, engineers, policy experts, and business leaders working together to build beneficial AI systems.
About the role
How do we know whether a model is safe - and how do we know whether the systems we built to catch misuse actually catch it?
Anthropic answers both questions with evaluations. We measure model behavior across misuse, prompt injection, and user well-being to inform training and deployment decisions. We also use AI to investigate potential misuse of Claude, analyzing real-world traffic to surface bad actors and emerging threats that drive enforcement actions. Neither is worth much unless the evaluations behind them are representative, robust, and trustworthy.
This role builds the methods and infrastructure that make them so. Sitting at the intersection of applied ML research and engineering, you'll design experiments to improve how we evaluate both model behavior and the agentic systems that govern it, build datasets that represent real abuse rather than clean approximations of it, and ship those methods into the pipelines that gate model training, agent changes, and launch decisions.
Responsibilities
Evaluation research and methodology. 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 measurement gaps, and evolve evals so they remain unsaturated and high-signal as model and agent capabilities advance.
Agentic investigation evals. Build and own the evaluation harness for an agentic investigation system — defining metrics, test cases, and grading approaches for a complex, long-horizon agent. Measure agent performance end-to-end (detection precision and recall, investigation quality, robustness) and drive hill-climbing on the hardest harm areas. Construct RL environments to improve Claude's safety investigation capabilities.
Datasets grounded in real harm. Construct high-quality eval datasets representing real-world misuse across harm areas such as cyber attacks, bio weapons, and influence operations, drawing from both real traffic patterns and synthetic generation. Collaborate with Policy and Enforcement to translate observed harm patterns into measurable evaluations.
Productionization and tooling. Ship successful research into evaluation, regression, and release pipelines that run during model training, on every agent change, prompt update, and underlying model upgrade, and beyond launch. Build tooling that enables policy experts to author, run, and iterate on evaluations without engineering support