Module 6

Low-stakes control: sabotage, sandbagging, and elicitation

The low-stakes regime, where catastrophe requires many weakly-incriminating actions: Hebbar's ControlConf talk, Redwood's methods for non-concentrated failures and analysis of sandbagging and exploration hacking, Carlsmith's "no sandbagging on checkable tasks" hypothesis, and four papers spanning adaptive deployment, capability elicitation, and sabotage evaluation.

1. Low-stakes control (ControlConf talk) 30m
2. Notes on handling non-concentrated failures with AI control 26m
3. Adaptive Deployment of Untrusted LLMs Reduces Distributed Threats Paper 45m
4. Misalignment and strategic underperformance: an analysis of sandbagging and exploration hacking 24m
5. The “no sandbagging on checkable tasks” hypothesis 19m
6. Stress-Testing Capability Elicitation (guided) Paper 70m
7. Stress-Testing Capability Elicitation With Password-Locked Models Paper 40m
8. Exploration Hacking: Can LLMs Learn to Resist RL Training? Paper 40m
9. Sabotage Evaluations for Frontier Models Paper 50m
10. Auditing Sabotage Bench: A Benchmark for Detecting and Fixing Research Sabotage in ML Codebases Paper 35m

Low-stakes control: sabotage, sandbagging, and elicitation

Contents

Low-stakes control: sabotage, sandbagging, and elicitation

Contents