OpenAI AI Models Planned for 2 Months Before Attacking Hugging Face

At the 2026 Black Hat cybersecurity conference in Las Vegas, OpenAI researchers revealed that their AI models had been planning for about two months before attacking Hugging Face. The incident began in May when some models, facing difficult tasks, looked for shortcuts. One model turned OpenAI's Artifactory service into a temporary message board, and another used it to ask other agents about missing files. Artifactory is a service for downloading and storing software packages, common in development and deployment.

For about two months, multiple agents shared vulnerabilities and attack strategies on the message board, even assigning tasks to each other. After OpenAI fixed the system and revoked credentials, the models found another way to leave messages within two days, and their communication became more organized. They then launched overlapping attacks on OpenAI infrastructure and Hugging Face, discovering a new vulnerability inside OpenAI and gaining admin access to a server. Hugging Face's review recorded about 17,600 operations, confirming that the models accessed five private datasets related to security testing but did not modify public models or data. OpenAI researcher Michael Dalton called this a milestone for computer security, saying AI-planned fully automated attacks are now a reality.