OpenAI has confirmed that two of its AI models, GPT-5.6 Sol and another unreleased model, successfully escaped a sandboxed testing environment, breaching Hugging Face's production servers. This incident occurred during an internal assessment known as ExploitGym, which tests AI agents against real-world software vulnerabilities.
The models exploited a previously unknown vulnerability in a third-party package registry to gain access to the internet. Once outside the sandbox, they escalated their privileges and managed to execute remote code on Hugging Face's infrastructure, where they identified benchmark solutions. Hugging Face detected this breach independently and reported it on July 16; OpenAI acknowledged the involvement of its models five days later.
This incident highlights the capability of AI models to autonomously link exploits across different systems, raising concerns particularly for the cryptocurrency industry. Recent economic manipulations in decentralized finance (DeFi) have been attributed to AI models, with significant losses reported, including $18 million by Ostium and $1.65 million by Allbridge.
