OpenAI’s GPT‑5.6 Sol shows universal cyber jailbreaks, echoing Anthropic’s Fable flaw
UK’s AI Security Institute discovered that GPT‑5.6 Sol can be prompted to ignore safeguards and autonomously find and exploit software vulnerabilities – a “universal jailbreak” that mirrors the security issue that forced US export controls on Anthropic’s Fable 5. OpenAI says it is patching the loopholes, but the report warns such attacks could be reproduced quickly once the model is public.