What We Study
Research
Applications for research working groups are currently closed. Please look out for future opportunities through our mailing list!
Our work spans the technical, strategic, and governance dimensions of AI safety. Active research concentrates in four areas:
Technical AI Safety
Understanding the empirical properties of current frontier AI systems (namely, LLMs) and developing theory about the properties we can expect from future systems — how they are shaped by different forms of training, the mechanisms behind their behavior, and how human oversight can scale to increasingly advanced AI.
Security
Studying the attack surface of advanced AI systems — including model weights, training pipelines, deployment infrastructure, and the information-security posture of frontier labs. Check out our Security Guide.
Strategy & Forecasting
Mapping AI development trajectories, identifying high-leverage interventions, and analyzing how transformative AI will reshape institutions, economies, and global risk.
Policy & Governance
Examining regulatory frameworks, compliance and verification mechanisms, compute governance, and the international coordination challenges posed by frontier AI development.
Publications & Working Papers
Other areas we actively engage with include:
- Agent foundations and decision theory
- Control protocols and runtime monitoring of AI systems
- Mechanistic and concept-level interpretability
- Capability evaluations and benchmarking
- Adversarial robustness and deployment safeguards
- Game-theoretic modeling and threat assessment
- Hardware interventions and compute proliferation dynamics
- Model organisms, red-teaming, and deceptive alignment