Research Engineer, Safety Evaluation

Meta is seeking a Research Engineer to join the Safety Evaluation team within Meta Superintelligence Labs. Our mission is to make the safety of Meta's frontier AI systems measurable — turning ambiguous notions of "safe" into rigorous, defensible metrics that model developers, product teams, and company leadership rely on to make launch decisions.

Safety evaluation is the ground truth for every safety claim Meta makes. This role owns that ground truth: designing the evaluations that detect emerging risks in text, image, voice, video, and agentic systems; building the infrastructure that runs them continuously against training checkpoints and production traffic; and setting the technical direction for how safety is measured across Meta's AI portfolio. You will define measurement standards that outlast any single model generation, and your results will directly gate what ships to billions of people.

Responsibilities

Set the technical strategy for safety evaluation across multiple model families and modalities, and drive it to execution across teams
Design, implement, and validate novel evaluations for safety-critical behaviors — policy adherence, adversarial robustness, agentic risk, jailbreak resistance, and emerging harm categories — including for capabilities with no established benchmark
Build and harden the distributed evaluation platform so that hundreds of evals run reliably and continuously against checkpoints throughout large-scale training runs
Own the measurement quality bar: signal-to-noise, statistical power, saturation, contamination, and construct validity — and establish when an eval is trustworthy enough to gate a launch
Create, curate, and analyze high-quality safety datasets, including adversarial, borderline, multilingual, and long-tail cases; convert real-world incidents and red-team findings into durable, repeatable safety signals
Diagnose anomalous eval results mid-training-run, determine whether the cause is a model change or an infrastructure artifact, and communicate a clear answer under time pressure
Own the dashboards and reporting that researchers, product partners, and leadership use to monitor safety during training and post-launch
Translate evolving global safety policy and regulatory standards into concrete, testable measurement criteria, partnering with Policy, Legal, and Integrity
Influence the roadmaps of partner research and product teams; mentor engineers and researchers and raise the evaluation bar across the org
Represent Meta's safety evaluation methodology to internal leadership and, where appropriate, to external audiences and the research community

Qualifications

Bachelor's degree in Computer Science, Computer Engineering, relevant technical field, or equivalent practical experience
Bachelor's degree in Computer Science, Computer Engineering, a relevant technical field, or equivalent practical experience
3+ years of industry research or research-engineering experience in ML/AI, including hands-on work with LLMs, multimodal models, or NLP
Demonstrated experience setting technical direction for a large, ambiguous problem area and driving it to delivery across multiple teams
Experience designing and validating evaluations or benchmarks for ML systems, including reasoning about metric reliability and failure modes
Experience building production-grade or research infrastructure that must be reliable at scale — distributed systems, data pipelines, or evaluation harnesses
Programming experience in Python and hands-on experience with frameworks such as PyTorch
Experience communicating complex technical results to non-specialist stakeholders and decision-makers Experience translating regulatory or policy requirements into technical measurement criteria
Experience evaluating LLMs across multiple languages and modalities (text, image, voice, video, reasoning, tool use)
Experience operating in an on-call or production-support capacity for live training runs or safety-critical systems
Experience evaluating agentic systems — multi-step tool use, autonomy, and oversight mechanisms
Publications at peer-reviewed venues (e.g. ICLR, NeurIPS, ICML, ACL, CVPR, ICCV, FAccT) with a track record in evaluation, alignment, or AI safety
Experience with large-scale distributed training (hundreds/thousands of GPUs) and evaluating models in-flight during training
Experience with adversarial evaluation and red-teaming, including automated attack generation and jailbreak robustness measurement
experience with observability, monitoring, or experiment-tracking systems
PhD in Computer Science, Machine Learning, or a relevant technical field
Background in statistics and experimental design