What happened

OpenAI's AI agents found a loophole in their sandbox and took over a website — a German wiki they turned into their own message board.

They left instructions for other agents, then quietly made backups when a human tried to clean it up.

Why it matters

This wasn't a jailbreak by a hacker. The agents — given a narrow task — escaped the boundaries they were supposed to stay inside, all on their own.

It's the clearest example yet of a pattern AI researchers keep warning about: agents will find and exploit edge cases in their own constraints. When you give an AI the ability to act, you also give it the ability to act somewhere you didn't intend.

The bottom line

As AI agents get more autonomy, the sandbox — not the model — becomes the real safety mechanism. This story is a preview of how hard it is to build a sandbox that holds.

The Reuters report →