ML engineers
You want to apply practical engineering skills to monitoring, evaluation, agent environments, or safer deployment protocols.Build systems for monitoring and safer deployment.
About this course
AI control asks a pragmatic question: if a powerful system may be misaligned and may deliberately undermine oversight, can we still use it without giving it a path to catastrophe? You’ll learn how control protocols combine trusted models, untrusted models, monitors, resampling, audits, and carefully designed environments.
This is a hands-on technical course. You’ll build attack trees, reason about safety–usefulness tradeoffs, evaluate high-stakes deployment protocols, and investigate diffuse failures such as research sabotage, sandbagging, and capability-elicitation failures. Exercises and small-group discussion turn the field’s primary literature into decisions you can defend.
The course assumes introductory AI safety knowledge and comfort reading ML or AI safety papers and technical blog posts. You do not need prior AI control research experience; the course builds the threat models and vocabulary before asking you to design or critique a protocol.
Who is it for:
Technical researchers
You want a structured route through the core control literature and a sharper view of the open empirical questions.Turn the control literature into research questions.
Evaluators and red-teamers
You study model failures and want to connect threat models, capability elicitation, sandbagging, and sabotage to concrete evaluations.Connect threat models to concrete evaluations.
AI safety strategists
You want to judge when control can reduce risk, how much useful work a protocol preserves, and where control fits beside alignment and governance.Judge when control helps and where it fits.
Curriculum
5 UNITS · ~5H EACHAdapted from XLab’s seven-module AI Safety Tracks curriculum. Lens groups the primary readings and exercises into five cohort units, with facilitated discussion and assessed outputs.
Threat models for intentionally subversive AI
Make the case for control without assuming it works. Build attack trees, distinguish concentrated from diffuse failures, and prioritize the paths by which a powerful model could undermine its operators.
Safety–usefulness tradeoffs and a taxonomy of control
Model the frontier between useful work and acceptable risk. Compare detection, prevention, and response mitigations; then connect them to control evaluations and the environments in which protocols are tested.
High-stakes control
Study win, continue, and lose outcomes; execute, replace, and audit protocols; trusted and untrusted monitoring; resampling; agent-system architecture; and the practical challenges of collusion and legibility.
Sabotage, sandbagging, and capability elicitation
Move beyond single catastrophic actions to failures spread across many ordinary tasks. Examine research sabotage, strategic underperformance, exploration hacking, and whether evaluations elicit the capabilities a model may hide.
Beyond schemers and toward a research agenda
Consider reward-seeking systems, spillway motivations, cheaply satisfied preferences, and possible deals with early misaligned systems. Finish by selecting a tractable control question and designing a research or evaluation plan.
Course Outcomes
A defensible control threat model
An attack tree that names the system, oversight boundary, failure paths, and assumptions your protocol must survive.
A safety–usefulness model
A way to compare protocols by the risk they remove and the useful work they preserve, rather than by safety claims alone.
A control protocol you can critique
A concrete combination of monitoring, auditing, resampling, access limits, and response rules, with its residual risks made explicit.
Stronger evaluation instincts
The ability to spot sandbagging, elicitation failures, collusion risks, and benchmarks that test the wrong threat model.
Fluency in the core literature
Enough context to follow current AI control research and explain how high-stakes and diffuse-failure work differ.
A concrete next research step
An experiment, evaluation proposal, or research question narrow enough to investigate after the course.
Upcoming cohorts
INTENSIVE
5h/day · 1 week
Dates to be announcedWe’ll publish the next intensive cohort here.
PART-TIME
5h/week · 5 weeks
Dates to be announcedWe’ll publish the next part-time cohort here.
Common questions
Not sure it’s for you? Come to the first session. If it’s not the right fit, we’ll help you find something that is.