AI Control

Hands-on technical track on controlling and evaluating AI systems. The introductory module works through Redwood Research's "AI Control: Improving Safety Despite Intentional Subversion" paper and then rebuilds its trusted-monitoring result as an interactive, model-backed demo.

6 modules

Module 1: Introduction, overview, and threat modeling
The foundation for the track: Redwood Research's "AI Control: Improving Safety Despite Intentional Subversion" paper rendered in full with guided exercises woven in, then rebuilt as an interactive, calibrated demo you can play with.
Start module Overview
Module 2: How useful is AI control?
The strongest arguments against AI control — deferred alignment research, slop, the superintelligence ceiling, moral hazard — illustrated one at a time, evaluated, and weighed against the additive alternatives; then the regime model: three states, a safety budget, and the loop that says when control is worth buying; and finally where control sits among Plans A through D, the plan landscape different levels of political will make available.
Start module Overview
Module 3: Areas of control work
Ryan Greenblatt's survey of the areas of control research and implementation that seem most promising, reproduced verbatim from the Redwood Research blog.
- An overview of areas of control work
Start module Overview
Module 4: Coming soon
Content for this module is in development.
Start module Overview
Module 5: High-stakes control
Opens with a talk on the high-stakes control roadmap, then three outline sections — what counts as high stakes, monitoring and mitigation, and environment construction — each with readings, from rogue-deployment threat models through control monitoring and Ctrl-Z's resampling protocols to why high-stakes research settings are hard to build. Closes with two decision-forcing walkthroughs of high-stakes control protocols, one argued and one measured: the untrusted advice protocol is the boundary case where a safety argument substitutes for a control evaluation; basic legibility protocols run a full red-team/blue-team control evaluation to settle whether letting an untrusted model comment its code helps the monitor catch backdoors. The section lessons are outline stubs while the full lessons are authored.
Start module Overview
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.
Start module Overview