Auditing Hidden Objectives

Describes a red team/blue team setup in which blinded auditors found a planted hidden objective, comparing eight techniques including SAE interpretability and training-data analysis. Teams evaluating models can adopt the audit structure.

• articleGlasswing Initial Update
• articleNatural Language Autoencoders
• articleProject Fetch Phase Two
• articleBuilding Ai Cyber Defenders

• articleAI models can provably hide arbitrary capabilitiessbulaev
• articleRed-Teaming Anthropic's Internal Agent Monitoring Systemsmetr.org
• articleAlignment FakingAnthropic research

Checking sign-in…

Loading comments…