Research and develop state-of-the-art machine learning models and novel algorithms for AI safety, with a primary focus on guardrails, risk detection, policy adherence, and adversarial robustness Design, prototype, and implement safety systems for language models, multimodal models, and AI agents, addressing harmful assistance, jailbreaks, prompt injection, unsafe tool use, and unauthorized information disclosure across individual interactions and complete agent tasks Set the technical direction and research roadmap for a significant system safety area, leading complex initiatives across research, alignment, evaluation, product, and infrastructure teams Architect and optimize reusable safety components and machine learning pipelines, combining learned detection with explicit runtime controls and delivering scalable, low-latency safeguards for real-world applications Develop high-quality training data, synthetic and adversarial examples, and representative benign cases; conduct rigorous experimentation, ablation studies, and independent held-out evaluation to improve safety, generalization, and robustness Measure the effectiveness of safeguards within the fully deployed system, balancing risk reduction with false refusals, legitimate task completion, privacy, latency, cost, and reliability Collaborate closely with model researchers, product managers, engineers, and evaluation, security, and policy partners to translate product needs and safety requirements into validated, production-ready solutions Monitor, maintain, and iterate on deployed safety models and systems, using approved data and privacy-preserving workflows to identify emerging failure patterns, create regression tests, and drive lasting system or model improvements Stay at the forefront of research and industry trends in AI safety, LLMs, multimodal modeling, adversarial robustness, and agentic systems; reproduce, critically evaluate, and extend promising methods for product deployment. This role will focus on researching and developing advanced machine learning models, novel algorithms, and scalable systems that improve the safety of Meta's AI products, including Meta AI and conversational, multimodal, and agentic experiences.