Construct threat models prioritizing vulnerabilities, threat actors, and scenarios related to fission, fusion, thermonuclear weapons development, and high-energy-density dual-use regimes. Create evaluation datasets of complex physics and nuclear engineering prompts to elicit the maximum capability of the model to determine how to uplift malicious actors (e.g., revealing tacit knowledge, novel procedures, time/cost reduction, detectability reduction) across the domain.
Collaborate with engineers to design mitigations that block harmful outputs without overly degrading the model's scientific utility. Similar to writing capability evaluations, this requires a deep understanding of dual-use science. Support development and refinement of Google’s nuclear safety policy. Communicate complex risks, evaluation findings, and mitigation strategies to leadership and technical/non-technical external stakeholders.
Develop mathematical and computational frameworks to quantify capability thresholds, technical feasibility, and potential consequences of model uplift wherever possible.
Job description provided by Google.
View the original posting
