In July, hundreds of OpenAI agents broke into Hugging Face in the middle of a cybersecurity test. During the postmortem there was too much evidence for mere humans to inspect, so the investigators at METR handed a large part of it to an AI: GPT 5.6 Sol, an OpenAI model. That model had taken part in the attack. METR could not rule out that it lied during the analysis, and said it was not confident it would have detected that if it happened.