Outside testing

In plain English

A lab testing its own model has a conflict of interest: it wants to ship. Outside testing brings in evaluators with no stake in the release, such as the UK AI Security Institute, the US CAISSI and independent organizations like METR and Apollo Research, to probe for dangerous capabilities and misaligned behavior before the public gets access.

The value depends on the details. Testers need enough time, access to the most capable version of the model and freedom to publish what they find. In practice access is often days or weeks, under agreements set by the lab, and results may be published only after company review. The Containment Board grades each lab on whether its latest flagship was tested by outsiders before release, how deep their access was and whether results were published. The EU General-Purpose AI Code of Practice now expects external evaluation of the most capable models, with limited exemptions.

Why it matters for rogue SI

Outside testers are often the first independent eyes on a new frontier model. Many early signs of rogue behavior come from their reports.

Laws and rules that use it

• Voluntary or guidance GPAI Code of Practice (Safety and Security) European Union
• Proposed FRONTIER Act United States
• Passed, not yet in force SB 813 California
• In force EO N-9-26 California
• Passed, not yet in force SB 315 Illinois
• Voluntary or guidance Voluntary Code Canada
• Voluntary or guidance UN Scientific Panel United Nations

Sources

1. The General-Purpose AI Code of Practice European Commission · Oct 7, 2026
2. H.R. 9925, FRONTIER Act, 119th Congress LegiScan