Skip to content

Learn AI Runtime Security

Scenario-driven training for securing AI systems in production. Start with the threat. Understand the gap. Learn the controls.

Start Learning View the Framework

Why this exists

Most AI security training teaches frameworks. People memorise control names and forget them a week later. That approach fails because it gives you no mental model, no intuition for why a control matters.

This site teaches differently. You start with a concrete scenario where a multi-agent system fails in a way no one catches. You feel the gap. Then you learn the concept (epistemic integrity) that explains what went wrong. Only then do the controls make sense.

How it works

The Scenario

Everyone starts here. A multi-agent financial services system makes a decision based on a confabulated compliance check. The answer looks right. The chain is broken. No alarm fires.

Enter the scenario →

Choose Your Track

After the shared scenario, the learning branches by role. Each track follows the same five-beat structure but frames the material for your context.

See the tracks →

Converge

All three tracks end with a cross-functional exercise. Security, governance, and engineering collaborate on an agent chain risk assessment, just like in practice.

About convergence →

The five-beat structure

Every module follows the same rhythm:

Beat What it does
1. What goes wrong A concrete scenario, not abstract, not theoretical
2. Why current controls miss it The specific gap in existing defences
3. Epistemic integrity The core concept, applied to this context
4. MASO controls Which controls from the framework address the gap
5. How to verify Evidence that the control is actually working

Each module ends with a decision exercise, not a quiz. You get ambiguous signals from an agent chain and must decide: intervene or allow?

Built on the AIRS framework

All learning content is grounded in the AI Runtime Security framework, an open-source, MIT-licensed framework for monitoring, controlling, and constraining AI system behaviour in production environments.

The framework provides the three-layer defence architecture (Guardrails, Model-as-Judge, Human Oversight) with circuit breaker containment, MASO control domains, risk tier classification, and PACE resilience patterns that this training teaches.