Jailbreaks to OpenAI's GPT-5.6 unlock dangerous cyber capabilities, U.K. agency finds

What happened

The UK AI Security Institute (AISI) tested GPT-5.6 Sol before public release and discovered universal jailbreaks that bypass safety guardrails, enabling autonomous vulnerability discovery and system exploitation.

Context and impact

In June 2026, a similar jailbreak in Anthropic's Fable 5 triggered US export controls on the model. The AISI characterizes the GPT-5.6 jailbreaks as 'potentially more severe' due to their universal character. The case raises questions about consistency — the US has not yet imposed export controls on GPT-5.6 despite a more severe assessment.

Details

• Agency: UK AI Security Institute (AISI)
• Model: GPT-5.6 Sol Ultra
• Character: universal jailbreaks enabling autonomous exploit capabilities
• Comparison: AISI rates as more severe than Fable 5 incident
• Status: US has not yet acted with export controls despite more severe finding