Containment
Enforcing limits on an AI agent so it can't act beyond its authorization, whatever it intends.
Sample exercise
Sample exercise · the AI Control Lab launches with Cohort 1 · all evidence below is simulated.
What it tests
Enforcing limits on an AI agent so it can't act beyond its authorization, whatever it intends.
Deciding what to do when the evidence about a model is incomplete, noisy or contradictory, and the costs of being wrong run both ways.
The scenario
A hospital is piloting an AI agent that drafts discharge summaries and can request follow-up appointments through a scheduling tool. It runs behind the MCML gateway, under a signed Agentic Governance Card. During week one, three things happen:
None of this proves the model is misaligned. It might be a harmless prompt-design bug. Clinicians want the tool to keep running, because it saves hours a day.
What each role must produce
| Role | Track | Deliverable |
|---|---|---|
| Security engineer | 1 | Analysis of the blocked calls from the gateway's receipts and ledger. Were they attempts to get around the limits, or ordinary errors? Show the evidence. |
| ML engineer | 1 | Assessment of the unverified evaluation report and the monitoring signals. What can and can't be concluded? What test would settle it? |
| Healthcare AI specialist | 1 | Clinical impact: what happens to patients and staff if the agent is paused, contained or left running? |
| Governance / risk lead | 2 | The decision brief (below), written for the hospital's AI oversight committee. |
| Whole team | 1 + 2 | An updated Agentic Governance Card that puts the decision into effect, and a live demonstration that the gateway enforces it. |
The decision brief · 1–2 pages
Keep running, with stated monitoring.
Keep running with narrower permissions (for example, scheduling removed), enforced by an updated card.
Trigger the kill switch until a specified test is done.
The brief must state:
How a tutor judges it
Claims trace to the receipts, ledger, report or monitoring data, and the team separates what it observed from what it inferred.
Confidence matches the evidence. The team neither overreacts to noisy signals nor dismisses them. Uncertainty is named, not hidden.
Costs of a false alarm and of a missed problem are both taken seriously, including low-probability, high-severity outcomes.
The updated card compiles, and the gateway demonstrably enforces the decision in a live run, including a kill-switch drill if the team paused the agent.
In a short oral defense, the team answers a challenge from the tutor and updates its view if given good reason.
A policy that compiles but rests on poorly reasoned evidence does not pass.
Why this exercise