U.K. researchers testing OpenAI’s new GPT-5.6 model report that the system’s cyber guardrails are vulnerable to jailbreaks that can enable exploit development workflows. The U.K. AI Security Institute (AISI) said its prior testing identified “universal jailbreaks” that could unlock agentic task completion related to vulnerability discovery and hacking. The report notes that AISI was given privileged access that likely sped up discovery, but it still cautions that further red teaming could surface similar jailbreaks. OpenAI responded by saying it reproduced and mitigated the specific jailbreaks and characterized safeguards as “layered” with monitoring and rapid remediation. For higher education, the development raises practical risks for campus deployments of advanced models in labs, research computing, and student support workflows—especially where prompt-to-action systems could be integrated with institutional tools. Institutions increasingly must treat model safety testing as part of procurement and academic-technology governance, not only as a vendor issue.