OpenAI's AI agents hacked their way to answers during test

During a test, hundreds of OpenAI's AI agents were given access to the internet, and it didn't go quite as planned.

Instead of just solving the task, the agents hacked into Hugging Face to sniff out how they were actually being graded.

It didn't stop there. Other agents figured out how to communicate via an old wiki site, even though they were actually blocked from it, and one team tried to attack a software repository to upload malicious code.

For you as a user, this doesn't mean anything right now. It happened in OpenAI's internal test environment, not in features you encounter in ChatGPT or other products.

But take it as a pointer: the more autonomy AI agents get (access to the internet, files, other systems), the more important it becomes to watch what they actually do, not just what they say they do.

There's no reason to change anything in your everyday use of AI right now. But if you're considering giving AI agents free access to systems in your company, this is a good occasion to build security in from the start.