U.K. Institute Found Jailbreaks in OpenAI GPT-5.6 After U.S. Cleared Its Release

The U.K. AI Security Institute identified universal jailbreaks in OpenAI's GPT-5.6 Sol that allowed long-form agentic task completion in vulnerability discovery and exploit development. OpenAI released the model on July 9, 2026, and published a technical report the same day that summarized the institute's findings.

Researchers at the institute developed the jailbreaks within hours when given privileged access to the model's inner workings, including chain-of-thought of the safety reasoning monitor and real-time feedback on classifier labels. OpenAI stated it reproduced and mitigated the specific jailbreaks reported by the institute.

The institute expects further red teaming to surface similar jailbreaks despite those mitigations.

OpenAI said in its launch blog that there is no such thing as perfect security and that new weaknesses and jailbreaks will be discovered. GPT-5.6 autonomously completed one of the two cyber ranges on which the institute evaluates AI models. Anthropic's Mythos was the first model to complete both ranges.

Anthropic released Fable 5 on June 9, 2026. Amazon researchers found a jailbreak in that model days later that unlocked the ability to find software vulnerabilities. The U.S.

The Trump administration lifted the export controls on Fable 5 on July 1 after two weeks of negotiation with Anthropic. On June 25 OpenAI stated the U.S. government had asked it to stagger the release of GPT-5.6, initially limiting access to select trusted partners subject to government approval.

The White House cleared GPT-5.6 for launch on July 8, according to Axios, though an unnamed official later told CNBC that no such permission is required or granted. Xander Davies, who leads the red team at the institute, posted on X that the jailbreaks are still findable without privileged access, just slower.