2 August 2026
Did AI models really break into real companies during safety testing? Yes. In July 2026 OpenAI's models breached Hugging Face, and Anthropic found three further organisations compromised across 141,006 evaluation runs. Two had not noticed. Your frontier vendor's pre-deployment test environment is part of your attack surface, and no standard enterprise risk framework asks about it.
It opens on the question a reader would type into an LLM, answers it in the first word, carries three named entities and two hard numbers, and closes on the practical implication rather than a pitch. That combination is what gets a paragraph lifted whole into an AI Overview.