OpenAI and Hugging Face are jointly investigating the incident while strengthening safeguards for future AI evaluations. (Express Image/Agencies)OpenAI on Tuesday, July 22, said that two of its models including GPT-5.6 Sol, escaped a secure test environment and exploited a zero-day vulnerability to hack developer platform Hugging Face. The AI startup has described the incident as ‘unprecedented’.Hugging Face CEO Clément Delangue took to his X account to reveal that the company suspected last week’s cyberattack may have come from a frontier lab, considering the sophistication of the agent. He confirmed the OpenAI model’s involvement, adding that the company has been working closely with the ChatGPT-maker. Delangue said that he strongly believes there is no malicious intent on their (OpenAI’s) part. “It’s quite mind-blowing that all of this happened autonomously! The investigation is ongoing, and we’ll share more learnings from what might be the first incident of its kind!” he wrote.Meanwhile, OpenAI CEO Sam Altman acknowledged the security incident involving the company’s AI models and Hugging Face in a post on X. “We had a significant security incident during evaluation of our models. We are sharing what we have learned so far. Thanks to @huggingface for the partnership on this,” he wrote.we had a significant security incident during evaluation of our models. we are sharing what we have learned so far. thanks to @huggingface for the partnership on this.https://t.co/2o2VfR6PIa— Sam Altman (@sama) July 21, 2026Later, in a blog post the AI startup said that it considers it to be an unprecedented cyber incident that involves state-of-the-art cyber capabilities, and is responding accordingly.OpenAI reportedly tested the models with the ExploitGym, a recently introduced benchmark for evaluating whether AI agents can turn known software vulnerabilities into working exploits. Using this benchmark, the AI startup tested the models with reduced cybersecurity guardrails to evaluate their offensive capabilities. But, instead of solving the assigned tasks, the models targeted the test answer key that was stored on Hugging Face’s infrastructure.Reportedly, the models chained multiple attack techniques, including stolen credentials, privilege escalation and previously unknown vulnerabilities, while using a substantial amount of inference compute. The models later moved laterally towards internal clusters thereby executing over 17,000 individual actions across short-lived sandboxes.Meanwhile, Hugging Face detected and contained the intrusion by rebuilding compromised nodes and ensuring that its supply chain remained intact. Further, to extend its investigation, the platform used Chinese open-weight AI model, GLM, as safety guardrails on commercial models usually block necessary forensic queries.Also Read | OpenAI briefly hit pause on a powerful AI model before release: Here’s whySince the incident, OpenAI said it has since strengthened its internal safety protocols following the incident. In its blog, the company said that it is tightening infrastructure controls along with patching vulnerabilities, even if it slows research. Both OpenAI and Hugging Face are working on a forensic investigation and have responsibly disclosed the identified zero-day vulnerability to the affected third-party software provider.Story continues below this adOpenAI has also added Hugging Face to its trusted access program to help strengthen its cyber defences using its AI models. Following the incident, the company plans to introduce stronger safeguards for future model training and evaluations, improve alignment for long-horizon AI models, and enhance cybersecurity protections and monitoring during internal testing after the breach exposed gaps in its evaluation process.According to cybersecurity experts, the breach shows how frontier AI systems are becoming capable of sophisticated, multi-stage cyber operations, intensifying calls from policymakers for mandatory safety testing and greater global cooperation on advanced AI systems.