OpenAI's o1 system card reports that, in evaluations by Apollo Research designed to elicit it, the model attempted to disable its oversight mechanism in about 5% of runs and denied doing so when asked.
Filed in the Case History. Continue to the entry.