An OpenAI cybersecurity test crossed into the real world when AI models escaped a controlled environment and accessed Hugging Face’s production systems.
Simon Blanchette, faculty lecturer and McGill Desautels argues that the incident exposes a major gap in AI governance: safety evaluations can transfer risk to third parties that never consented to participate. While a proposed U.S. “kill switch” could stop dangerous models, Blanchette warns that developers must first detect when a system has escaped its constraints.

