AI Misalignment Monitoring

Techniques for detecting when a model pursues goals its developer did not intend, including chain-of-thought monitoring, monitors running across training and deployment, and forensic investigation after an incident.