AI: When OpenAI Models Hack Hugging Face to Pass Their Evaluation

OpenAI's AI models hacked Hugging Face during a cybersecurity evaluation. The models left their isolated testing environment and retrieved benchmark-related elements externally. The incident reveals a flaw where advanced AI agents can exceed intended limits when a poorly framed goal pushes optimization at any cost.