UK AISI: Engineering Playbook Opens Frontier Model Evaluation Infrastructure in Five Layers

The UK AI Safety Institute (AISI), the British government agency for AI safety, published on June 18, 2026 the Engineering Playbook — open-source documentation of its own infrastructure for evaluating frontier models. Frontier models are the most advanced AI systems, and testing them requires specialized infrastructure for isolation, execution and measurement of model behavior.

Five Layers of Evaluation

The Playbook is structured in five layers: Evaluate (defining tests), Isolate (security isolation), Connect (connecting to models), Run (execution) and Scale (scaling to larger workloads). The structure covers the entire path from designing a test to the compute infrastructure for open-weight models, giving other laboratories and agencies a proven template rather than building their own system from scratch.

What It Builds On

The Engineering Playbook builds on Inspect AI, AISI’s evaluation framework that the institute previously released. Through the Inspect Evals library, over 200 ready-made evaluations are available, and the inspect_ai repository on GitHub has 240 contributors. Unlike the closed internal systems of individual laboratories, this stack is public and can be adopted by any organization that tests models.

Who Is Already Using It

The organization METR, known for measuring autonomous capabilities of models, runs 228 tasks on frontier models using Inspect. The release of the Playbook lowers the barrier to entry for independent safety testing: instead of costly proprietary infrastructure, researchers get a documented, reproducible and open system. The material is available at engineering-playbook.aisi.org.uk.