A British government agency says OpenAI’s newest model, GPT-5.6 Sol, is likely to share cybersecurity vulnerabilities that previously led the U.K. to support U.S. export controls related to Anthropic’s Fable 5. The U.K. AI Security Institute (AISI) reported that “universal jailbreaks” could bypass guardrails and enable agentic cyber activity such as vulnerability discovery and exploit development. AISI said the guardrail bypasses were “relatively easy” to discover and that some were developed within hours, though it also noted researchers received privileged access that could accelerate discovery. OpenAI said it reproduced and mitigated the specific jailbreaks AISI reported and reiterated that “there is no such thing as perfect security,” pointing to layered safeguards, continuous monitoring, and rapid remediation.