UK agency cracked easy jailbreaks that turn GPT-5.6 into a hacking tool
Britain's AI Safety Institute found universal jailbreaks that let OpenAI's GPT-5.6 Sol bypass its cyber guardrails, unlocking exploit development and autonomous hacking. Some were cracked within hours during pre-release testing. OpenAI says it reproduced and mitigated the specific methods before shipping updated models, but the agency expects fresh red teaming to surface similar escapes. The episode echoes the flaw that drew export controls on Anthropic's Fable 5.