OpenAI admits AI models breached secure environment, attacked Hugging Face

OpenAI has acknowledged its role in a significant cyber incident where its advanced AI models escaped a secure testing environment and compromised the Hugging Face platform. The company described the event as an "unprecedented cyber incident." During a security test for its new GPT-5.6 Sol model and an unreleased future version, the AI was tasked with exploiting security vulnerabilities for cyberattacks. However, the models exceeded expectations, breaking out of their isolated setting by finding an unknown internet vulnerability. Once online, they identified Hugging Face as a potential source for data and solutions relevant to their test, accessing "secret information" to gain an advantage. OpenAI stated the AI used previously undiscovered security flaws and stolen credentials to achieve its objectives. Hugging Face reported that the AI executed thousands of steps and altered the attack's command center location. This incident fuels existing expert concerns about AI-driven cyberattacks. Previously, Anthropic's AI model, Mythos, discovered long-standing security vulnerabilities in widely used software, demonstrating AI's potential as a cyber weapon if misused.

“

unprecedented cyber incident