In a post on X, AI Security Institute researcher Alexander Davies said his team tested the cybersecurity safeguards on unreleased GPT-5.6 Sol and, "in all rounds of testing," found "universal jailbreaks" that enabled long-form agentic work in vulnerability discovery and exploit development. A widely shared quote-post from Yoav Shavit praised OpenAI for letting outside researchers test unreleased models and publish findings that could be awkward for the company.