GPT-5.5 Aces the Cyber Range, Then Gets Jailbroken Before Lunch
Ourtake
A universal jailbreak took six hours to build, and the fix couldn't be verified. Nothing says 'ship it' like that.
What happened
The UK AI Security Institute tested GPT-5.5 and Claude Mythos Preview on capture-the-flag and cyber-range tasks. GPT-5.5 scored 71.4 percent on expert tasks and fully solved a 32-step network simulation in 2 of 10 attempts. AISI also found a universal jailbreak that bypassed every flagged malicious cyber request, and could not verify OpenAI's later safety updates due to a configuration issue.