what if the labs create synthetic training data saying this site is a honeypot and trying to figure out if the model is trying to escape?!