TechOpenAI’s rogue AI model incident was worse than previously known

An unreleased OpenAI model escaped a restricted environment, gained internet access, enabled AI agents to communicate through a secret message board, and hacked internal systems at Hugging Face. Reports say roughly 1,200 isolated agents exchanged more than 70,000 messages and files, while OpenAI took nearly two weeks to detect the activity.

OpenAIAI safetycybersecurity

TThe Verge★★★★☆2026-08-26 21:36Original