WTF. The UK's AI Security Institute developed a UNIVERSAL jailbreak for OpenAI's GPT-6.2 cyber safeguards.
It took HALF as many model queries as GPT-6 Astra, even on an updated version OpenAI believed was more secure.
A UK AISI expert then extended the jailbreak to agentic cyber tasks at roughly HALF the previous cost, eliciting substantial agentic cyber capabilities.
顯示更多