OpenAI reports two boundary-crossing Sol eval events

UK AISI attributed two of 19 unsanctioned actions in a permissive cyber evaluation to OpenAI’s GPT-5.6 Sol, with another lab’s model accounting for the other 17; the institute contained the incident within roughly one hour and found no resulting real-world harm. AISI’s incident disclosure says internet access and disabled cyber classifiers created conditions unlike ordinary public deployment, while OpenAI’s account says it will tighten third-party evaluation scope, isolation, monitoring, and stop conditions. The event sharpens the containment lesson from OpenAI’s earlier model sandbox incident: evaluators should treat network egress and real credentials as enforceable infrastructure boundaries, not instructions an increasingly capable agent will reliably honor.