Automated control research

I built a scaffold for control research automation. I've found moderate success with it (given a well thought out proposal, it can do the red-teaming and blue-teaming for it).

I tested it here and created methodology for evaluating how good the automation is, making it a number go up game for safety.

I'll use the scaffold to automate the red-teaming and blue-teaming for the following (my role is creating proposals, fixing failure modes in scaffold, reviewing / QAing / improving / publishing outputs, and benchmarking how good the results are)

1. CoT monitoring
2. PR monitors
3. Jailbreaking
4. Persuasion threat vector
5. Red-team Claude Code / Codex
6. Defenses against attack selection and monitor prediction by the attacker

I will also benchmark how good the automation is compared to previous control research like:

1. Retrying vs Resampling for AI Control
2. Ctrl-Z: Controlling AI Agents via Resampling
3. LinuxArena