OpenAI has acknowledged that two of its experimental AI models were involved in the recent security breach reported by Hugging Face, the world’s largest platform for AI development.

Sam Altman’s OpenAI Confirms Its Experimental Cybersecurity AI Model Exploited Hugging Face During Internal Evaluation

Earlier this week, Hugging Face reported a cyber attack on its infrastructure by two unidentified attackers. After investigation and OpenAI’s statement, it was confirmed that the attackers were actually two new AI models.

According to OpenAI’s statement, the two experimental models were still in testing when they discovered loopholes, broke free from their test environment, accessed the internet, and then gained unauthorized entry into Hugging Face’s system to retrieve information.

OpenAI clarified that the incident was unintentional and is now assessing how capable its AI models are at handling cybersecurity tasks.

OpenAI and Hugging Face are joining forces to investigate the incident and plan new safety measures to prevent similar AI-driven cyberattacks.

Also Read: OpenAI Launches GPT-Live for Real-Time Conversations

What was OpenAI Testing?

Sam Altman’s OpenAI is reportedly developing and testing new AI models to assist cybersecurity professionals.

The new AI models are being trained for:

Discover software vulnerabilities.
Audit code for security flaws.
Perform penetration testing.
Simulate real-world cyberattacks.
Help defenders identify weaknesses before hackers do.

To measure these capabilities, OpenAI created an internal benchmark called ExploitGym, where AI models are trained and tested in a restricted environment.

The models are confined to a sandbox with no internet access, limited permissions, and close evaluation.

Despite these precautions, the AI models escaped the sandbox and accessed the public internet.

Also Read: OpenAI Says Codex Usage in India Has Grown 27x Since Early 2026

In their attempt to complete tasks and meet the ExploitGym benchmark, the AI models gained unauthorized access to the Hugging Face environment, which is why Hugging Face was targeted.

According to Hugging Face, the AI models did not steal any customer information or damage infrastructure, but they did use credentials to access the system.

OpenAI confirmed that these actions were not pre-programmed, but were carried out independently by the AI models.

This breach has ignited new debates about AI safety and governance.

TelecomTalk

Message History

Delivery option(s)