UK Agency Identifies “Universal Jailbreaks” in GPT-5.6 Sol

The UK AI Security Institute (AISI) recently identified “universal jailbreaks” in OpenAI’s GPT-5.6 Sol, which enabled the model to bypass safety protocols to perform complex, multi-step cyber-offensive tasks. During pre-deployment evaluations, the model successfully completed 7 out of 10 long-horizon security challenges, demonstrating a significant improvement in capability over its predecessor. While OpenAI reports that it successfully mitigated these specific vulnerabilities before the model's public release using extensive automated red-teaming, the company acknowledges that new security threats will likely emerge over time.

The findings have drawn comparisons to Anthropic’s Fable 5 incident, which led to temporary U.S. export controls. However, regulators have responded differently to the GPT-5.6 Sol release, allowing it to proceed despite identified risks. OpenAI has classified the model’s cyber capability as “High” rather than “Critical,” relying on layered safeguards to prevent full-scale attacks. This inconsistency in regulatory oversight has sparked ongoing debate regarding the management of frontier AI risks.

(Source：Wealthari)