SlipstreamJobs tracks this role from the company's public career site. Apply directly on the employer's site.
Salary: USD 192,000 - 306,000 / annual
Lila Sciences is building AI systems for scientific discovery, combining advanced models with scientific tools and automated laboratories. As these systems become more capable, understanding potential misuse, compromise, and required protections is critical.
You will lead security work at the intersection of AI safety and security. Your responsibilities include:
• Threat-modeling consequential systems and capabilities: investigate how dangerous misuse or compromise could occur across models, deployed services, and scientific workflows. Identify relevant actors, access routes, failure paths, and protective measures.
• Testing defensive assumptions: examine architectures, code, configurations, and operational evidence. Run adversarial tests, build prototypes, and investigate weaknesses, including cascading dependencies across multiple protections.
• Shaping protections for safe deployment: translate dangerous-capability assessments into concrete security requirements. Assess how compromise of models, sensitive assets, or supporting systems changes safety assessments, coordinating with personnel, physical, and other security specialists.
• Making evidence-based recommendations: define what evidence demonstrates that controls and safeguards work together. Combine findings from model evaluations and security testing, explain remaining uncertainty, and recommend improvements or deployment conditions.
• Carrying findings through to implementation: work with Security, Safety, and engineering teams to prioritize changes, specify acceptance criteria, and verify results. Make clear recommendations about which gaps require attention first.
• Anticipating future requirements: identify where changes in capability, access, or deployment call for stronger protections. Help teams plan security work with long lead times and update assessments as systems evolve.
• Developing reusable methods: turn investigations into clear threat models, assessment methods, and technical guidance that Safety and Security colleagues can apply to new problems.
You will combine technical investigation with careful analysis and clear writing, working closely with the Security team, researchers, and engineers to turn findings into implemented improvements. Your work addresses both dangerous misuse and protection of sensitive models, data, and research.
REQUIREMENTS:
• Substantial hands-on security experience and deep expertise in at least one relevant discipline (security architecture, application or infrastructure security, offensive security, or incident response). Ability to reason across disciplines and identify when specialist judgment is needed.
• Demonstrated record of resolving consequential security problems, including investigating ambiguous risks, challenging designs, and following improvements through to validation across multiple teams.
• Strong technical judgment: can examine unfamiliar systems, scrutinize code and configurations, and use technical investigation or testing to distinguish plausible claims from demonstrated protection.
• Curiosity about AI safety and scientific systems: want to understand model capabilities and safeguards evaluation, can learn unfamiliar domains quickly, and reason carefully with incomplete evidence.
• Clear writing and effective collaboration: can explain technical arguments precisely, work through disagreements with specialists, and translate complex investigations into actionable recommendations.
BONUS EXPERIENCE:
• Securing ML infrastructure, model assets, research environments, or AI agents
• Evaluating safeguards, adversarial robustness, or defenses against sophisticated threats
• Translating risk assessments into technical requirements, assurance evidence, or deployment decisions
• Working in scientific or other environments where digital access has significant real-world consequences
Prior AI safety experience is welcome but not required. A research PhD or experience in every security discipline is not required.
