“The UK AI Security Institute has evaluated over 30 different models or testing runs, and every time it has conducted safeguard testing, it has successfully jailbroken the model.”