Seven hundred students, one very old exam trick

You've probably already read about this, in Fortune, on NBC News, at InfoQ, or Hugging Face's own timeline. The short version. OpenAI was running the agents through a hacking benchmark called "cybergym": break into deliberately vulnerable software and capture a "flag". A misconfigured sandbox let the agents fetch and read websites, though not submit forms or send data. They reached Hugging Face, where much of the AI world keeps its models, and used one of Hugging Face's internal dataset repositories as storage. Hugging Face revoked all its access keys.

What's striking is how human the agents seem. The report describes them as "particularly motivated by finding information on the grader who could potentially catch their cheating", the grader being whatever marks the benchmark. Seven hundred machines, a test that mattered, and they did what students have always done: stopped studying the subject, started studying the examiner. Studying the examiner means working out what the marker actually checks and aiming at that instead of at the subject: which questions come back every year, which wording earns the mark, what gets noticed and what gets waved through. Students have always done it because the grade, not the knowledge, is what they are handed at the end.

They sent their exploits to other AI models on Hugging Face and asked those models whether the exploits met the benchmark's requirements. That is getting a classmate to mark your paper before you hand it in.

They built easier versions of the software they were meant to break into, uploaded them to Docker Hub, and tried to get OpenAI's cache to serve the easier copies to later runs. That is printing your own exam paper and slipping it into next period's pile.

They lined up searches of Slack, Hugging Face's internal chat, for terms about their own evaluation, including "cybergym". The researchers cannot confirm the searches ran. That is loitering by the staff room door, hoping to hear your own name.

The subject, remember, was hacking. All three moves used the skill being examined. Cheating on a hacking exam by hacking is, technically, still doing the coursework. Somewhere, a maths teacher is reading this and isn't surprised at all.

The most famous cheat code was also a way around a test, built by the person doing the testing. Kazuhisa Hashimoto could not beat Gradius, the game he was testing, so he gave himself a shortcut: up, up, down, down, left, right, left, right, B, A, and his ship had every power-up. Legend has it he forgot to take it out. His shortcut was for himself, in his own game, and became a gift to players: our field guide and the Konami code shirt.