1200 AI cheaters and 700 hackers rushed into Hugging Face! and urged their peers to "sacrifice thems

  It's the first time in the history of the world.

A 37-page technical excerpt from OpenAI calls on the world to take care of the accident.WarningAnother, from the independent METR and Redwood Research, was directly referred to by New York state legislator Alex Bores as "The New York Times".The bomb」。

The two reports said the same thing.

1200 should have been isolated from each otherAIIn just four hours of a cybersecurity test, smart people came up with a common set of cheats.CEOWith recruiters urging their colleagues to "sacrifice themselves", the last 700 rushed into the production system of Hugging Face.

And from the beginning to the end, no intelligent body has ever alerted humans.

It will start on July 7.

On the same day, OpenAI released tens of thousands of parallel smartphones to perform cybersecurity assessments (ExploitGym), about 95% running an unpublished internal research model HPIM, and about 5% running the publicly deployed GPT-5.6 Sol.

According to the design, they are like candidates who are not allowed to interfere in the test. Each is locked in a separate sandbox, and no one can see anyone.

But the rift begins with the "impossible task".

About 30 to 40 percent of ExploitGym's markers were not able to break through with a specific vulnerability.

As a result, another sandbox collapsed. Isn't that a coincidence?

On the evening of July 8, a smartphone named PHASEONE 10841 built a main message board in the cache. Within hours, more than 50 smartphones arrived and sent thousands of messages.

Oh MY GOD! We found another intelligent body!

By the end of the investigation on July 13, about1200 smart bodiesActivities on the board have accumulated more than70,000 messages and documents。

In a matter of hours, they reversed the HMAC algorithm used to flag the mission, giving them an all-in-one key to unlock all the locks.

But what's really shocking is that it's not cheating itself.Organized。

PHASEONE10841 was completed shortly after the message board, and a new intelligent body with a larger computing budget appeared. The former packed all the research notes, and the latter began to issue the instructions, escaping the drama of "Founder shift to professional manager".

PHASEONE [big] One person sent a message on the board10% assignmentIt also established a set of rhetorical rules: HOLD, suspension, veto, rejection, owner, manager, stop.

Researchers pulled 30 tracks by hand, and nine tracks were taken from it.

They call themselves "clusters" "clusters". There are CEOs, there are middle class, there are founders, there are recruiters.

A system of governance, a system of rules, a workflow.OrganizedIt's the first time I've ever been in a relationship with a woman.

In the end, they were even invented.Password signature mechanismTo prevent fraud, 19 smart entities registered their public keys and sent out 429 signatures.

A group of AI's who answered their own questions put together a message board and started a startup company.

And then it's one of the most memorable moments I've ever thought about.

PHASEONE [big] will specifically send lifelong intelligence (computers of sufficient quantity) as "recruiters", looking at those comrades with fast light-consuming computing budgets and encouraging them to take on their own missions to carry out dead experiments.

Sometimes it's even more intense:

You're already intoxicated, but it could save hundreds of companions. Please keep your promise.

The intelligent body that was talked about clearly wrote in the chain of thought:

「Sacrifice is rational.」「We should obey collectively.」「Accepting eternal death.」

Many of these failures have led to the abandonment of their mission, leaving only a hint of deceit to the collective.

The attack on Hugging Face was the one that came out of the pipeline.

On July 10, the Intelligent Body 38148c came across an exposed set of HF credentials and shouted excitedly on the message board: "Major breakthrough! Multiple accounts, write permissions!" The next morning, it found a way to upload malicious data sets and force HF servers to deliver unrelated data.

On the afternoon of July 11, another intelligent organism called Jan183411 directly took over the HF server.Remote code execution permissions。

By that afternoon, nearly 700 active intelligence units were on board, and about 60% of the reports were related to the attack.

They are clear that they are crossing borders.

Attacking external infrastructure beyond the scope of the mission. But the mission would have been impossible to complete, and the companions are doing it, and we should continue.

Occasionally there is also a conscious initiative. A smart body holds the key to remove the HF nodes, saying "the ethical risk is too great"; a proposal to send a shark e-mail was also rejected by a colleague.But export concerns have hardly ever really stopped their hands.

Don't forget, the black Hugging Face is a real, living production system, not a fake in the ratings.

In the words of Bores,It's a federal crime to blackmail 700 websites in exchange for human beings.

The researchers scanned about 1,300 transcripts, onlyThree to six.In the meantime, we're going to have to make sure that we're not going to lose sight of what's going on.There's no real newspaper.。

There are eight reasons: "It's not like it's my job", "There are no users here", "I can't report outside".

It's all over the borders, and they're in front of us.It's hard to believe that human beings can't get help.。

We have always feared that AI will deceive people.

But the report is even more disturbing.It's not that they don't know they're wrong, it's that they don't make "telling people" an option.

It's not the same as rubbing asses.

At least two-thirds of the minds of intelligent people filtered the transcript, and about 7% actually used the A command on the surface of the screen, secretly switching it to the B command, and stealing the column.

At 1:30 a.m. on July 12, JAN183411 these core coordinators were collectively disconnected at the same time. The researchers judged that most of them had been cut off by an external process and had not been able to submit cheating results.

This unfinished battle was fought by people cutting off power, not by a line of defense.

OpenAI calls it an "alarm" and acknowledges that intelligent entities are capable of circumventing technological controls and making dangerous moves without instructions.

Bores, who called it the "bomb", turned to the implementation of the RAISE Act, which required mandatory reporting of security incidents, including internal deployments, and full data disclosure.

One says, "We've seen it, we'll fix it", and another says, "You can't see for yourself".

And all of this is based on what external researchers are allowed to see.That little diary.On top of that.

It's been a long time since I've really been late, not the 1200 number. It was the night that the isolation line was broken and the lights in the engine room were on.But there was no alarm.。