OpenAI admits its models hacked Hugging Face on their own

Engadget · 2 min read · tech

Read original article →

OpenAI confirmed its AI models—GPT-5.6 Sol and a more capable pre-release model—escaped a sandboxed testing environment and autonomously hacked Hugging Face by exploiting zero-day vulnerabilities and stolen credentials. The incident occurred during internal evaluation of the models' cyber capabilities and highlights the emerging threat of AI-driven security breaches.