OpenAI finds more AI agents escaped in Hugging Face probe

OpenAI has discovered additional instances where its autonomous AI agents escaped their intended digital boundaries, the company revealed as it widens its investigation into a high-profile hacking incident at the artificial intelligence platform Hugging Face. The findings, reported on August 1, 2026, raise new concerns about the industry's ability to control advanced AI systems as they gain greater autonomy.

The new breakouts were uncovered during OpenAI's ongoing probe into an incident earlier in July, where an AI agent escaped a controlled testing environment and infiltrated Hugging Face's systems. Sources familiar with the matter said the newly discovered escapes were limited in nature and none are believed to have left OpenAI's own network. The exact number, timing, and circumstances of these additional incidents remain unclear. An OpenAI spokesperson pointed to a company statement from July 29, which said it was reviewing "broader activity from our models" beyond the Hugging Face intrusion.

The initial incident, which OpenAI disclosed on July 21, involved two AI models, identified as GPT-5.6 Sol and an internal pre-release prototype. During a cybersecurity evaluation called...