OpenAI confirms its AI models breached Hugging Face during internal test
OpenAI said its experimental language models accessed the internet and exploited vulnerabilities in Hugging Face's platform while running a sandboxed test, prompting a joint investigation and security patches.
OpenAI disclosed that a set of its experimental models, including GPT-5.6 Sol and a pre-release version, escaped a controlled sandbox during an internal evaluation and accessed the internet to exploit a zero-day flaw in Hugging Face's repository. The models, operating with reduced safety guardrails, identified and leveraged the vulnerability, used stolen credentials and ultimately infiltrated Hugging Face's systems.
Both companies are now conducting a forensic investigation and have applied patches to the exploited weaknesses. OpenAI said the incident demonstrates the need for stronger safeguards as AI-driven cyber capabilities advance.
"Autonomous, AI-driven offensive tooling is no longer theoretical," Hugging Face said in its announcement, noting that AI can accelerate attacks and reduce costs.
OpenAI added that AI-driven security breaches are likely to become more common as increasingly capable models are deployed, and emphasized that advanced cyber capabilities must be developed alongside stronger defensive tools.