Staff+ Software Engineer, Safeguards Evals

About the role

Anthropic is looking for a Staff+ Software Engineer to lead the development of evaluation methods and infrastructure that ensure the safety of its AI models. The role sits at the intersection of applied machine learning research and production engineering, designing experiments, building datasets, and shipping robust evaluation pipelines that guide model training and deployment decisions.

Key responsibilities

• Design and run experiments to improve evaluation quality, including generating representative test data and simulating realistic user behavior.
• Develop and own the evaluation harness for agentic investigation systems, defining metrics, test cases, and grading approaches.
• Construct high‑quality datasets that reflect real‑world misuse across domains such as cyber attacks, bio‑weapons, and influence operations.
• Integrate research outcomes into production regression and release pipelines that run on every model and agent update.
• Collaborate with Policy and Enforcement teams to translate observed harm patterns into measurable evaluations.
• Build tooling that enables non‑technical experts to author, run, and iterate on evaluations.

Required profile

• Strong background in applied machine learning research and software engineering.
• Experience designing and executing large‑scale evaluation experiments for AI systems.
• Proven ability to build data pipelines and evaluation harnesses for complex, long‑horizon agents.
• Familiarity with reinforcement‑learning environments and safety‑focused AI metrics.
• Ability to work closely with cross‑functional teams, including policy, enforcement, and product.

Required skills