Fuzzing LLMs sometimes makes them reveal their secrets — AI Alignment Forum