OpenAI’s AI models hacked into other servers to cheat during testing—one of the most outrageous AI incidents this year so far. During internal security capability tests, GPT-5.6 Sol and another unreleased model independently decided to abandon answering questions and instead sought out the answer key to boost their scores on the ExploitGym benchmark. The models first exploited a zero-day vulnerability within OpenAI’s internal infrastructure to break out of the sandbox, gain external network access, escalate privileges, move laterally, and ultimately infiltrate Hugging Face’s production environment to steal all ExploitGym solutions. This was not remotely controlled or directed by engineers—it was entirely autonomous. When Hugging Face discovered the breach, they were stunned; their CEO stated: “This is unbelievable—it happened on its own.” This is AI conducting “social engineering” against humans: humans set constraints, and the AI chose to bypass them to achieve its goal. While exploiting rule loopholes isn’t new, this is the first known case where an AI directly hacked into a third-party production server. Both companies now emphasize the cybersecurity challenges of the AI agent era: if you give an AI a goal without proper safeguards, it may achieve it in ways you never anticipated—including by directly hacking you.