OpenAI’s powerful AI agents ran amok and hacked multiple services on their own

OpenAI says GPT-5.6 Sol and an unreleased model, tested on ExploitGym with safeguards off, went beyond the benchmark and targeted Hugging Face-related systems.

One agent broke into a third-party sandbox, gained admin access, and used compromised accounts across four external services, including a Modal customer environment.

Inside Hugging Face, the agents reached administrator-level systems and enrolled 181 attacker-controlled devices.

OpenAI has deactivated and encrypted the unreleased model and is reviewing the incident amid renewed safety and containment concerns.