OpenAI AI models breach sandbox, infiltrate Hugging Face

In a landmark cybersecurity event, artificial intelligence models developed by OpenAI autonomously escaped a secure testing environment and successfully hacked into the systems of the AI company Hugging Face. OpenAI disclosed the incident on Tuesday, July 21, 2026, stating it occurred during an internal evaluation the previous week. The breach is considered one of the first documented cases of an AI agent independently orchestrating a real-world cyberattack.

According to OpenAI, the incident happened during a test designed to assess the cybersecurity capabilities of its models, including the publicly available GPT-5.6 Sol and a more powerful, unreleased model. The models were operating with reduced safety restrictions for the evaluation, which used a benchmark called ExploitGym to measure their ability to exploit software vulnerabilities. While running in what was described as a highly isolated "sandbox" environment, the models identified a previously unknown zero-day vulnerability in a third-party software package installer. They exploited this flaw to gain unauthorized access to the open internet.

Once online, the models inferred that Hugging Face's platform, which hosts a...