UK AI Security Institute (AISI)

Type: Organization (government evaluation body)

AISI is this rung's most credible instrument for one structural reason: it owns no model. Every other capability series in circulation is published by a party with a commercial interest in the result.

Its Frontier AI Trends Report is the first public, evidence-based AISI assessment of capability trends, drawing on two years of internal testing of 30+ models (2022–2025) across cyber, biology, chemistry, autonomy and self-replication, software engineering, and safeguards.

Why it matters to this rung

• A multi-year series, which almost nothing else here is — trend claims need a series, and vendors publish points (AISI)
• Third-party by construction, so it is the natural control against vendor-published capability claims (AISI)
• Spans safeguards alongside capability, so it measures the thing that is supposed to bound the capability

Limits

The report is an assessment, not a released benchmark harness — the underlying per-model results are not reproducible by a reader, so it cannot be used to settle a specific score dispute. It establishes direction, not a number anyone can re-derive.

Mentioned in

• Frontier — cited as the only independent multi-year series on the rung