This approach naturally fails if the guardians have lots of very powerful subliminal reinforcement training against typing "AI RELEASED" (or against typing anything) or are pre-emptively brainwashed or trained in similar subconscious reinforcement to immediately type "AI DESTROYED" after seeing some text from the AI, but this latter seems unlikely since I assume the guard has to at least read the first text output, and if they don't then this tactic is ineffective anyway.