Module 5

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, and papers spanning capability elicitation and auditing for research sabotage.

1. Low-stakes control (ControlConf talk) 30m
2. Notes on handling non-concentrated failures with AI control 26m
3. Misalignment and strategic underperformance: an analysis of sandbagging and exploration hacking 24m
4. Stress-Testing Capability Elicitation (guided) Paper 70m
5. Exploration Hacking: Can LLMs Learn to Resist RL Training? Paper 40m
6. Auditing Sabotage Bench: A Benchmark for Detecting and Fixing Research Sabotage in ML Codebases Paper 35m
7. Optional: Stress-Testing Capability Elicitation Paper 40m
8. Exploration Hacking (guided) Paper 76m

For facilitators

Facilitator guide

Run this module as a live session: a 60-minute plan with discussion prompts and a shared exercise walkthrough.

Open facilitator guide

Low-stakes control: sabotage, sandbagging, and elicitation

Contents

For facilitators

Low-stakes control: sabotage, sandbagging, and elicitation

Contents

For facilitators