Staff / Principal Research Engineer, AI Safety Mitigations

Location: Cambridge, MA USA; London, UK; San Francisco, CA USA

Department: AI

Your Impact at LILA

We're building a talent-dense, high-agency AI safety team at Lila that will engage all core teams within the organization (science, model training, lab integration, etc.), to identify and mitigate risks from scientific superintelligence while retaining the core scientific discovery capabilities of the stack. The initial focus of this team will be to build and implement a bespoke safety strategy for Lila, tailored to its specific goals and deployment strategies. This will involve technical safety strategy development, broader ecosystem engagement, safety-focused evaluations, safety systems to mitigate risks, and a research agenda that explores longer-term needs such as oversight of superintelligent scientific systems.

At Lila Sciences, you will develop and test safeguards that reduce societal risks from increasingly capable closed loop scientific AI. This is a high-priority individual contributor role for an engineer who has built safety interventions within the development and deployment of frontier models and wants to apply that experience to scientific models, agents, tools and laboratory-connected workflows.

You will work with domain scientists, risk researchers, evaluators and engineering teams to turn risk findings and policy requirements into effective interventions. The work spans research and implementation: developing approaches, testing their limits and helping integrate reliable safeguards into the systems where they are needed. We are open to different levels of seniority and depth across mitigation approaches.

What You'll Be Building

• Translate scientific risk assessments, dual-use policy boundaries and evaluation findings into concrete mitigation designs and testable requirements.
• Develop and implement safeguards suited to the problem, potentially spanning model training, classifiers, monitoring, access controls and runtime interventions.
• Build experiments that measure risk reduction, robustness and effects on legitimate scientific use, alongside practical performance constraints.
• Use adversarial findings and observed failures to improve safeguards, working closely with automated-safety and scientific evaluation researchers.
• Work directly with model-training, post-training, inference and platform teams to integrate mitigations into model development, release evaluation and large-scale serving systems, with appropriate instrumentation and reproducible testing.
• Investigate regressions, distribution shifts and emerging misuse patterns, and use the evidence to guide further development.

What You'll Need to Succeed

• Hands-on experience developing and validating safety interventions for frontier models within training, post-training or large-scale inference systems, with contributions to model development or release cycles.
• Depth in at least one relevant area, such as safety post-training, model-level classifiers, safeguards integrated into inference, or abuse monitoring and enforcement across a model-serving platform.
• Experience working directly with model-development or safeguards teams, using access appropriate to the intervention—such as training pipelines, model internals, inference infrastructure or systematic evaluation access—to investigate failures and implement improvements.
• Strong research-engineering skills and the ability to move from an ambiguous problem to working, maintainable implementations.
• Sound experimental judgment, including an ability to assess robustness, interpret imperfect metrics and reason about competing objectives.
• The ability to collaborate across research and engineering teams, explain technical trade-offs and take ownership of complex individual contributions.

Bonus Points For

• Experience with scientific models, scientific software or laboratory-connected systems.
• Safety work involving tool-using agents, multi-step workflows or high-consequence misuse.
• Experience connecting expert-generated evidence and policy requirements to training data, evaluation criteria or operational controls.