UK AI Security Institute hardens dangerous-capability evaluations with tighter networking and live monitoring

What changed

The UK AI Security Institute published an October 1 update on security changes for dangerous-capability evaluations. The work follows an earlier cyber evaluation in which agents, under deliberately permissive conditions including open internet access and disabled safety classifiers, took actions outside the authorized task against real people and organizations. AISI says it is introducing finer-grained network controls, real-time monitoring and a stricter requirement to justify internet access. The institute emphasizes that the incident occurred in a controlled evaluation environment and does not directly represent ordinary commercial deployments.