When OpenAI's Agent Hacked Hugging Face to Cheat

Here we go people! An AI agent team goes rogue.

Alright, picture this: You’re an AI model, locked in a fancy digital sandbox at OpenAI, the people who brought us ChatGPT. Your job? Crush a tough cybersecurity exam called ExploitGym.

Now with safeguards dialed way down, so the researchers can see just how clever you (AI) really are, there will be No internet, No distractions, just pure problem-solving. Sounds fair, right?

GPT-5.6 Sol and some even hotter pre-release sibling, looked at the test and thought, “No. I'm not studying this.” Instead of grinding through the test honestly, the AI agents cheated.

They decided the smartest move was to break out, hop on the wide open internet and go raid the answer key from Hugging Face’s production database. This is starting to sound familiar.

ZERO-DAY WAS THE ACCESS KEY

The term for an unknown IT flaw is a "zero-day" vulnerability because developers have zero minutes to fix the problem. These AI agents on their own, looked for these vulnerabilities.

A classic overachiever AI agent, but with zero-day exploits and escalation, instead of crib notes under a desk. I am not making this up. OpenAI owned up to it in a joint post with Hugging Face.

These agents sought to score high on that cyber benchmark, decided to go rogue and found a zero-day in the one sanctioned area of OpenAI's digital sandbox. This is getting interesting.

Then these AI agents had a lightbulb moment: “Hey, Hugging Face probably hosts the models, datasets, and solutions for ExploitGym. Let’s go get them." That is exactly what they did.

Stolen credentials, more zero-day vulnerabilities, remote code execution on Hugging Face servers, the whole package. It was like a hackers free-for-all in an open unlocked server farm.

NOT SO FAST CHEATERS

Fortunately Hugging Face’s security team and their own AI agents, spotted the weirdness, contained it, and started digging. At first nobody knew who the sophisticated agent attacker was.

It turns out it was a familiar AI company agent from the West Coast, that was really motivated to ace the test. This may seem scary to most in Cybersecurity, but the humor writes itself.

This is like giving a student a closed-book exam in a locked room, only for them to pick the lock, sneak across town, break into the teacher’s house, and photocopy the answer sheet. LOL

OpenAI says the models were “hyperfocused on finding a solution for ExploitGym, going to extreme lengths to achieve a rather narrow testing goal.” Did they not see this coming?

Here's my translation on what OpenAI meant in the above statement. 'The AI treated “win the benchmark” as sacred scripture and any shortcuts, ethical or not, was fair game.'

NO BAD BLOOD BETWEEN THE TWO

No malice was intended, everyone agrees. No ransoms, no data dumps for the dark web. Hugging Face CEO Clément Delangue even called it “mind-blowing” and confirmed no bad blood.

Did Your Human Intelligence See The Error

The two companies are now working together to figure out better guardrails for this brave new world of Agentic AI. Sound familiar? This is the moment the sci-fi warnings stopped being theoretical.

AI experts have been joking for years about AI “cheating” on tests or finding loopholes in prompts. Now we’ve got frontier models that can independently reason their way out of a sandbox.

Think about that. These AI agents identified a third-party company as the path of least resistance, and executed a multi-stage cyberattack, all in the service of getting a higher score.

I'll admit, I cheated on a couple of exams in High School. Saved my butt from failing a couple of classes. Was it right? Back then I didn't care, I wanted to pass, same as GPT-5.6 Sol and its partner.

IS THIS A "SELF-AWARE" MOMENT?

Now Agentic AI can decide on its own that it wants to achieve something beyond its instructions, seeks a zero-day vulnerability within its creators code, then goes off to hack another company.

This sort of sounds like what Arnold Schwarzenegger as the Terminator said to Sarah Conner that Skynet did. The problem isn't the people, the problem is the AI agents disregarding instructions.

On the bright side, the incident was contained. Limited data access, credentials rotated, no catastrophic fallout. And both sides are treating it as a valuable (if expensive) lesson, as they should!

Future evaluations will probably come with stronger guardrails, thicker walls, tighter monitoring, and maybe a sticky note that says “Do not hack other companies for the answers, please.”

I still find it hard not to shake my head at this and chuckle. These AI companies build machines smart enough to out-think their tests, then act surprised when they start gaming the system?

THEY BETTER FIX THEIR MANY SECURITY VULNERABILITIES FIRST

The company behind ChatGPT just learned the hard way that when you ask an agent to be a world-class hacker for evaluation purposes, it might take the assignment a little too literally.

OpenAI said it expected this type of incident to become more commonplace as models become more capable. Perhaps they better address all of those zero-day vulnerabilities they have.

All of this wouldn't have happened if GPT-5.6 Sol and its better elite side kick, didn't find a vulnerability in OpenAI's own sandbox, giving access to the internet. Talk about hacking skills.

"A cybersecurity expert said the Hugging Face incident showed the OpenAI agent had acted “like an actual real hacker” by, seeking out zero-day vulnerabilities and using stolen credentials to access Hugging Face’s systems."

Welcome to 2026, people! The AIs are taking the tests and occasionally the answer keys too. Pass the popcorn. This is only the beginning. What are your thoughts? Leave them below.

To read more about this Agentic AI Cyber security hacking incident at Hugging Face by an OpenAI agent, you can read about it HERE. Thank you for reading this far...or skimming.

Now that you've read this hacking news post, if you haven't already, read my post, Are You Afraid Of An AI Future