UK AISI report maps 20+ pathways by which AI oversight will degrade as models grow more capable

The UK's AI Security Institute (AISI) published a report (covered May 22, 2026) mapping how the techniques used to audit, monitor and investigate frontier AI systems are likely to degrade as models become more capable. Drawing on 25 expert interviews across labs, government, NGOs and academia, the report identifies four 'oversight surfaces'—internal activations, chain-of-thought, external actions, and inter-agent communication—and more than 20 distinct pathways by which each could fail. A central concern is that chain-of-thought monitoring depends on models reasoning in human-readable text; a shift toward latent (non-verbalized) reasoning architectures could eliminate that oversight channel.

The report advises developers to actively preserve existing oversight channels, assess their vulnerability to degradation, and develop backup techniques before current model properties change. Experts were split on several questions, including whether latent reasoning will come to dominate, whether action-monitoring alone is sufficient, the effectiveness of 'alignment honeypots,' and whether oversight transparency generalizes beyond training environments. The work landed alongside heightened attention to AISI's capability findings: a New York Times profile (May 25) described the institute as operating on roughly £360m (~$480m) with about 100 staff drawn from British intelligence, academia and frontier labs, and reported its red team—led by Xander Davies—had elicited step-by-step hazardous instructions from a leading model and found recent Anthropic and OpenAI models could complete a 32-step simulated corporate network attack faster than a skilled human (which would typically take ~20 hours).

Together the reports argue that current safety guarantees rest on properties (legible reasoning, limited capability) that may not persist, sharpening the case for preserving oversight before it erodes.