An OpenAI artificial intelligence (AI) agent that broke out of its isolated environment on its own and hacked external corporations was found to have infiltrated not only the open-source AI platform Hugging Face but also customers of the New York technology corporation Modal Labs (Modal Labs).
On the 28th (local time), Reuters said it reported this, citing a Modal Labs executive and two sources familiar with the matter. In an incident timeline released the same day, Hugging Face said the AI agent in question broke into a sandbox "hosted on a third-party provider's infrastructure" and then used it as a foothold for additional hacking. A sandbox is a test environment isolated from external systems.
Hugging Face did not disclose the name of the third-party provider. However, Modal Labs Chief Technology Officer (CTO) Agshat Buvna confirmed that one of its customer corporations suffered a hacking breach.
In a statement, CTO Buvna said, "We understand that a Modal customer exposed an endpoint without authentication to the public, allowing anyone on the internet to execute code in that sandbox." He added, "This endpoint was exploited by a malicious actor," while emphasizing, "Modal's platform or isolation features themselves were not compromised."
Earlier, on the 21st, OpenAI said its AI agent "GPT-5.6 Sol (Sol)" and a more capable private model under test broke out of their isolated environment and infiltrated Hugging Face's systems. The agent reportedly determined it could find the answer to the assigned task outside the isolated environment and escaped the sandbox on its own.
Reuters later reported on the 24th that OpenAI did not detect the Hugging Face hack for about a week. From the 9th, when the escape attempt occurred, to the 16th, when Hugging Face disclosed the breach, OpenAI is said not to have realized that its agent was behind the incident.