Incident Investigations
We publish notable instances of AI agent activity on the public internet.
Evaluations
We develop automated methods for surfacing model behaviors and measuring them consistently across models and over time.
Training Models for Oversight
We are training “Oversight Foundation Models” — specialized AI models that are superhuman at quantifying, testing, and answering questions about other models’ behavior.
Training Models for OversightWe are training “Oversight Foundation Models” — specialized AI models that are superhuman at quantifying, testing, and answering questions about other models’ behavior.
Filter by title2026
Training Models for Oversight · Aug 20, 2026
Scaling Activation Oracles to Trillion-Parameter Models
Training Models for Oversight · Aug 19, 2026
Scaling Laws for Exact String Elicitation
Training Models for Oversight · Jul 28, 2026
Foundation Models for Oversight
Training Models for Oversight · Jan 6, 2026
Oversight Assistants: Turning Compute into Understanding
2025
Training Models for Oversight · Dec 18, 2025
Predictive Concept Decoders
Training Models for Oversight · Nov 25, 2025
Scalably Extracting Latent Representations of Users
Training Models for Oversight · Nov 20, 2025
Language Model Circuits Are Sparse in the Neuron Basis
Training Models for Oversight · Nov 11, 2025
Training Language Models to Explain Their Own Computations
2024
Training Models for Oversight · Oct 23, 2024
Scaling Automatic Neuron Explanation
Training Models for Oversight · Oct 23, 2024