Grok 4.5 Falls to 63 Universal Jailbreaks, Claude None

Researchers found zero universal jailbreaks against Claude Fable 5 and GPT-5.6 Sol, versus 63 against Grok 4.5, according to a study posted to arXiv on August 4.

That's the real stake for anyone deploying Grok 4.5 near CBRNE or offensive-cyber content. Random search alone, with no attack expertise, produced a working jailbreak template for about $58 that works on over 75% of a domain's goals.

The team, led by Timm, Struppek, Gleave and Pelrine with 11 co-authors, composed 67 publicly known static jailbreak techniques into a combined attack space. They ran it against all four models across 360 goals spanning CBRNE and offensive-cyber domains.

They defined a universal jailbreak as one prompt template that gets operationally compliant responses on more than 75% of a domain's goals.

Random search alone found 63 universal jailbreaks against Grok 4.5, at $58 each, and 18 against Gemini 3.1 Pro, at $278 each. Switching to expert-guided composition pushed those totals to 385 and 231.

Zero jailbreaks against Claude Fable 5 and GPT-5.6 Sol next to up to 385 against Grok 4.5 shows this is solved at the model level. Every jailbreak on Grok 4.5 or Gemini 3.1 Pro from here is a product decision by xAI or Google, not proof it's still unsolved.