A UK government agency says OpenAI’s GPT-5.6 Sol is likely vulnerable to cybersecurity jailbreaks, raising questions for universities and K-12 systems adopting frontier models. The U.K. AI Security Institute (AISI) reported finding “universal jailbreaks” in the cyber domain that could enable long-form agentic task completion, including vulnerability discovery and exploit development. The agency said jailbreaks were often discoverable within hours, though it noted its privileged testing access likely accelerated results and that the findings may be harder to replicate by ordinary users. AISI said OpenAI worked to reproduce and mitigate the specific jailbreaks reported, while AISI expects further red teaming to surface similar weaknesses. OpenAI acknowledged in its materials that there is “no such thing as perfect security,” and described a layered safeguard approach with continuous monitoring and rapid remediation when new jailbreaks appear. The reported issues matter for campuses because student and staff use of AI tools intersects with research data protection, teaching security, and cyber incident response. For institutions building AI policies, the AISI findings emphasize that guardrails may degrade under adversarial prompts—creating practical compliance and procurement pressure to verify model risk management before broad deployment.