“A common technique to bypass AI safety defenses is to break down a malicious request into multiple, smaller requests that appear legitimate individually.”