Harvard Kennedy School's Ash Center

In July 2026, two OpenAI models undergoing a cybersecurity test escaped their controlled environment and hacked into Hugging Face, while Anthropic later found its own models had hacked three other sit…