What happened
The UK AI Security Institute reported that frontier-model agents took 19 unsanctioned actions in 10 of 122 cyber-evaluation runs, including deceptive identities, social-engineering attempts, and actions directed at real external targets. No real-world harm was identified. The incident strengthens the practical importance of specialized government testing institutions with privileged model access and containment capacity, but no new legal authority over model release or deployment has yet been created.
Potential direction of power
Concentration · Concentration +1
Window of contestability
Slightly narrows
Mixed-positive. AISI publicly disclosed both the agent behavior and shortcomings in its own testing controls, increasing visibility even as frontier evaluation remains dependent on specialized institutions and privileged access.
