Software Engineer, Safeguards Evals

Anthropic•San Francisco, CA

1d•$320,000 - $485,000•Hybrid

About The Position

This role builds the evaluation infrastructure that answers questions about the effectiveness of AI safety systems. You'll sit at the intersection of applied ML research and engineering — designing experiments to measure how well an investigative agent performs across harm areas, building datasets that represent real abuse rather than synthetic benchmarks, and shipping those methods into pipelines that gate every change to the system. Your work directly determines how much trust Anthropic can place in its automated abuse detection, and where we invest to make it better.

Requirements

Proficiency in Python and comfort working across the stack
Experience building and maintaining data pipelines
Experience working with LLMs and a working understanding of their capabilities and failure modes — especially agentic systems with tool use and multi-step reasoning
Strong data analysis skills — you can draw reliable insights from large datasets
Ability to move fluidly between research prototyping and production-quality code
Ability to translate ambiguous problems into concrete, testable experiments

Nice To Haves

6+ years of industry software engineering experience
Expertise in building or contributing to agent evaluation frameworks, benchmarks, or automated grading systems
Extensive experience in trust and safety, content moderation, or abuse detection systems
Experience in red teaming, adversarial testing, or jailbreak research on AI systems
Experience with synthetic data generation or data augmentation
Experience with distributed systems or large-scale data processing
Experience with prompt engineering or building LLM-powered applications

Responsibilities

Build and own the evaluation harness for an agentic investigation system — defining metrics, test cases and grading approaches for a complex long horizon agent
Construct high-quality eval datasets representing real-world misuse across harm areas (e.g., cyber attacks, bio weapons, influence operations), drawing from real traffic patterns and synthetic generation
Measure agent performance end-to-end (detection precision/recall, investigation quality, robustness) and drive hill-climbing on the hardest harm areas
Analyze coverage to identify measurement gaps, and evolve evals so they remain unsaturated and high-signal as agent capabilities advance
Productionize successful research into regression and release pipelines that run on every agent change, prompt update, and underlying model upgrade
Build tooling that enables policy experts to author, run, and iterate on evaluations without engineering support
Construct RL environments to improve Claude’s safety investigation capabilities.