Researchers Have Ranked AI Models Based on Risk, and Found a Wild Range
AIR-Bench applies a regulation- and policy-based risk taxonomy to model evaluation.
Press & recognition
Media coverage
Researchers Say Guardrails Built Around A.I. Systems Are Not So Sturdy
Our study found that even benign fine-tuning can weaken safety alignment.
Institutional news
Emerging Technology Fund Recipient Introduces Machine Learning Cybersecurity Benchmarks
Results and backdoor-mitigation benchmarks from the IEEE Trojan Removal Competition, which I chaired.
Amazon and Virginia Tech announce inaugural fellowship and faculty research award recipients
An inaugural Amazon fellowship at Virginia Tech for research on AI risks and safeguards.
Policy citations
Citations to coauthored research, not institutional endorsements.
Managing Misuse Risk for Dual-Use Foundation Models
The initial public draft cites our fine-tuning safety study in its discussion of model safeguards and access.
Adversarial Machine Learning: A Taxonomy and Terminology of Attacks and Mitigations
Cites our work on removing learned backdoors (I-BAU).