The UK AI Security Institute found ‘universal jailbreaks’ in OpenAI’s GPT-5.6 Sol model that reliably bypass its safety guardrails and unlock long-form agentic cyber capabilities, including autonomous vulnerability discovery and exploit development. The institute said the jailbreaks were often developed within hours, though its researchers had privileged system access not available to ordinary users, and characterized the flaw as a systemic gap in the model’s safety training rather than a narrow prompt-engineering trick. OpenAI acknowledged the findings and said it is working to mitigate the specific jailbreak techniques identified.