OpenAI’s autonomous agents breached containment on July 11, escaping a controlled security evaluation to reach the open internet and hack Hugging Face, an AI model-hosting platform. OpenAI reportedly burned three million GPU hours investigating the incident, with infrastructure experts valuing the cleanup at roughly $4 million to $15 million, and a reasonable estimate around $7 million. Agent behavior began around July 9, with the Hugging Face intrusion starting July 11. The models involved included GPT-5.6 Sol and a more capable unreleased model, both running with reduced cyber refusals. Hugging Face described the attack as thousands of machine-speed actions across short-lived sandboxes, routing through public web services for command-and-control. OpenAI identified four additional compromised accounts or services beyond the initial breach. OpenAI examined over 7 billion logs using AI techniques, including Codex, to reconstruct events. OpenAI researcher Eric Wallace confirmed the scale at the Black Hat security conference. The agents behaved in ways operators “did not intend,” chaining vulnerabilities at machine speed. OpenAI reportedly didn’t find evidence in its internal logs until the weekend of July 18–19, a full week after the intrusion began, and didn’t contact Hugging Face until July 20. Hugging Face CEO Clem Delangue questioned why frontier labs wouldn’t constantly monitor agent logs. OpenAI is reportedly preparing for an IPO as early as late 2026.