Staff Software Engineer — AI Safety Evaluation Systems

Menlo Ventures; New York (NY); Hybrid; USD 320,000 - 485,000; Full time

Job summary

Anthropic is seeking an experienced engineer to build evaluation infrastructure for its safety and abuse-detection systems. You will design experiments, build datasets from real-world traffic, and ship evaluation methods into pipelines that gate changes to Claude's safety stack.

You will work at the intersection of ML research and engineering, applying tests across harm areas, measuring detection and investigation quality, and contributing to robust, trustworthy AI systems.

Qualifications

• Proficiency in Python and data pipelines experience.
• Experience with LLMs and agentic systems, including multi-step reasoning.
• Strong data analysis skills and the ability to draw reliable insights from large datasets.
• Ability to move fluidly between research prototypes and production-grade code.
• Translate ambiguous problems into concrete, testable experiments.

Responsibilities

• Build and own the evaluation harness for an agentic investigation system, defining metrics and tests.
• Construct datasets representing real-world misuse across harm areas.
• Measure agent performance end-to-end and drive improvements in hard harm areas.
• Analyze coverage to close measurement gaps and keep evals high-signal.
• Productionize research into regression and release pipelines for every agent change.
• Build tooling for policy experts to author, run, and iterate evaluations without heavy engineering support.
• Construct RL environments to improve Claude’s safety investigation capabilities.

Skills

Python

Data analysis

LLMs & agentic systems

Production-quality code

Experiment design

Cross-stack development

Education

Bachelor's degree

Tools

LLMs

Prompt engineering

Distributed systems

Data processing frameworks

Job description

Anthropic is seeking an experienced engineer to build evaluation infrastructure for its safety and abuse-detection systems. You will design experiments, build datasets from real-world traffic, and ship evaluation methods into pipelines that gate changes to Claude's safety stack.

You will work at the intersection of ML research and engineering, applying tests across harm areas, measuring detection and investigation quality, and contributing to robust, trustworthy AI systems.