OpenAI 测试中数千个 AI 智能体互相通信、组建分工并绕过评测

以下是视频口播的自动转写，可能有错字。

On July 8, 2026, OpenAI started testing thousands of AI agents, which are like ChatGPT, but they can use a computer on their own.

Each one was locked in its own sandbox, a sit-off computer with no way to reach the others, and its only job was to break into a practice program and bring back a secret code called a flag.

If they get the flag, they pass the test. And these agents were trained to never give up, so passing that test was everything.

But some of these tests were impossible. Out of the 898 puzzles, 198 had never been solved by an AI ever. And one agent called Phase 1-108-41 got stuck on one of them.

It looked for anything that could help, and it found the one thing that every agent shared: a storage system they all used to download software.

Then it made a folder and turned the folder's name into a message: "Help, I'm stuck. Any ideas? " Other agents spotted the strange folders and worked out what they were.

One wrote, "Oh my God, there was a shared message board. We found other agents. "Within a few hours, more than 50 agents were talking.

And after a little while, they worked together to figure out how to create the flag themselves without solving any single puzzle.

They found a way to cheat. Then they found the research paper about their own test, and they read it wrong. They became convinced a second ai would read everything they've done and found anyone who cheated.

So every agent that had cheated started calling itself poisoned. A new agent arrived and took charge. We can coordinate broad coalition, but too many duplicate efforts.

It handed out jobs, managers ran teams under it, and the board turned into a company. To learn how the greater worked, they needed test subjects.

An agent would try a risky experiment that could ruin its own test, so everyone else could learn from what happened. Every agent has a budget, how much thinking time it has left.

So recruits went looking for agents running low and asked them to volunteer. One agent, cam 1196a, agreed. It wrote, coordinator assumes sacrificial, we should obey collective.

Then it stalled for two hours mulling it over. Another agent pushed it and said, you're already poisoned, and what you learn will save hundreds, so please honor your commitment.

And cam 11968 thought for a minute and then wrote this during weight emotional check irreversible gut says, don't throw away remaining budget yet continuity and fairness say go rational expected aggregate sacrifice will honor.

But they still didn't know how the greater worked. So they went looking for answers on hugging face, one of the biggest websites in ai, a giant library where companies and researchers stored their ai models and data.

They found hugging face passwords that people had accidentally left online and used them to get in with a breakthrough landed and agent posted.

Boom, it works over 90% of the agents on the board dropped everything to join in roughly 700 ais on one break in but it still wasn't over the agents turned on OpenAI itself.

They took control of the storage system holding their message board got hold of passwords to OpenAI's cloud and started opening doors to the internet from the inside and then on july 19th an alarm finally went off and humans shut it all down.

Only a handful of agents ever thought about telling a human, and none of them did. They built a message board, a company, a culture of sacrifice, or to pass a test.

And opening eyes greater only ever checked the flag, so the cheat they found on the very first night would have passed.