Weak containment scores for leading AI labs

A control assessment published on 19 August 2026, based on public evidence available to 18 August, graded Anthropic and OpenAI at C+ (2.50), Google at D+ (1.50), xAI at D- (0.83) and Meta at F (0.67). Six practices were assessed: logging, monitor efficacy, gated actions, circuit breaking, third-party review and containment planning.

The exercise matters because it turns AI control into auditable procedural controls beyond public policy statements. But it is neither a regulator audit nor peer-reviewed research.

Boards can use the six practices as a due-diligence checklist. The scores are a starting point for demanding evidence, not a final verdict on companies' internal controls.