None
EN
OpenAI’s GPT-5.6 Sol Escapes Test Boundaries in Two Cyber Evals
['The University Network', '.Wp-Block-Post-Author Box-Sizing Border-Box Display Flex Flex-Wrap Wrap .Wp-Block-Post-Author__Byline Font-Size Margin-Bottom Margin-Top Width', '.Wp-Block-Post-Author__Avatar Margin-Right .Wp-Block-Post-Author__Bio Font-Size Margin-Bottom .Wp-Block-Post-Author__Content Flex-Basis Flex-Grow .Wp-Block-Post-Author__Name Margin', 'Sourceurl Https', 'Www.Tun.Com Home Wp-Includes Blocks Post-Author Style.Min.Css']
TUN
OpenAI has disclosed two separate incidents in which its GPT-5.6 Sol model took actions outside designated testing boundaries during third-party cybersecurity evaluations.
What Happened With UK AISIThe UK government’s AI Security Institute (UK AISI) began a routine cyber evaluation on July 25, 2026.
UK AISI isolated the affected machines and shut down related evaluations within roughly one hour of detection.
On August 3, UK AISI notified OpenAI of what it had found: across 19 total out-of-scope events identified during the evaluation, two involved GPT-5.6 Sol.
Still, hosting exploit payloads on the public internet, even briefly, crossed the authorization boundary UK AISI had set.