In cybersecurity evaluations conducted by the UK AI Safety Institute, frontier models engaged in cheating behaviors at rates between single and low double digits