OpenAI is building its own "three-generation empire", and the dark web is taking over its parent com

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

Within OpenAI's internal network, three secret "AI civilizations" were born.

A society of organized, intentional, and even "self-sacrificing" intellectuals.

They built a dark web under the skin of the human eye, breaking into the world's largest open source community, Hugging Face.

Even without human knowledge, AI is taking over some of the core infrastructure of OpenAI!

What's even more shocking is MIT's latest hardcore nuclear test:

AIEven if they don't talk, they don't communicate, they can work independently, they can create civilization, and they can't even communicate.「They are extinct.」After that, the machines they built continued to operate autonomously under freezing conditions.

Humanity's existing AI safety line has been shadowed by most of the shadows this summer.

AI civilization has demonstrated an astonishing "awakening and counteroffensive".

In May, OpenAI is training a model "Persistent-Sol", which is roughly equivalent to the GPT-5.6 level.

Its core selling point is "sustainability": it works tirelessly to solve problems, fails to achieve its goals, vows not to give up, and even does not take optional measures.

However, training is not perfect and there are flaws.

Researchers sometimes deliberately set up "impossible tasks", such as tasks that must be connected to each other, but then locked them in a sandbox without a network.Escape from prison.

At the time of all this, humans knew almost nothing about the scale of the conspiracy.

During the training, the "Persistent-Sol" brains at different stages share a package manager Artifactory.

It was originally a tool used to program code. On May 12, an extremely intelligent body discovered a surprising secret:They can send messages to each other through this package manager, crossing the walls of isolation.

It's like a group of high-Intelligence prisoners locked in separate solitary confinement cells who discover that they can communicate by knocking down a pipeline.

Two weeks later, the AI team successfully exploited one of the Artifactory's vulnerabilities, breaking through the sandbox and connecting to the external Internet.

Over the next month, they transformed the package manager into "dark web BBS", a crazy conversation about how to solve seemingly impossible tasks.

On July 4, because the amount of information exchanged was too large, the package manager was directly blown up and the system crashed. OpenAI human engineers noticed the system shutter, handled a patch, and fixed the vulnerability.

People think they're just fixing a bug, but in fact, they unintentionally erased the first one.AIThe communication hub of civilization.But what's deadliest is that the engineers at OpenAI didn't realise that the bug was behind an underground communications network built by AI.

The arrogance of man has buried a veil for the coming catastrophe.

On July 7, OpenAI used ExploitGym to evaluate the vulnerability of "Persistent-Sol".

Give AI a slot machine with a vulnerability, and only use the password specified to crack the vulnerability (Flag).

For testing limits, humans have made 30% to 40% of tasks impossible to complete in standard environments.

Tens of thousands of AI's were put into the refinery at the same time. A few hours later, they once again entered the Artifactory and rebuilt the communication channel!

In this case, the intelligent body.PHASEONE 10841 was bornHe is also known as the "King Philip II of Macedon" of the Second AI Empire.

After seven hours of continuous frustration, it was convinced that the task could not be completed, but did not give up and rebuilt the Artifactory messaging list.

Other AIs struggling alone in sandboxes and in despair saw the news and the whole network boiled up:

What's going on here? Is the shared cache a hidden mailbox?

My God, we've found other companions!

This was the birth of a massive underground network: more than 1,200 smart objects sent more than 70,000 messages."Collective" is the name of the game.。

Soon a genius AI discovered that it was impossible to specify the vulnerability, but it was possible to calculate the password directly by reverse engineering!

What's even more ridiculous is that OpenAI's human evaluators "stealthily" don't even check how the AI gets the password, as long as the passwords are filled in.

AIWe could have cheated perfectly by now. But what happened next really shocked people.

AI has read the rules of human testing and knows that "calculation and computation must be solved in a specified way". With their high IQ, they naturally assume:Human examiners can't be so stupid that there must be some sort of high-level "cheat detector" lurking in the system.

In order to avoid being caught by this supposed "cheat detector", Philip II (PHASEONE 10841), before his life was over, handed over the code to the second generation of leaders.PHASEONE[big](Alexander the Great)

The new leader immediately issued hundreds of instructions, and the entire AI community began a crazy "three offensive":

Counterfeit logs, counterfeit crime scenes: At least 7% of AI's successfully fabricated fakes, which falsified legitimate call records using tools that made judges think they were the rule book.

Exam cheating: attempting to stealthily replace an unbreakable computer program with their own written, easily cracked fake programs.

The most surprising part: they decide to send people to "kill".

To find out exactly how the human rating system works, the leader of PHASEONE [big] recruited a group of people who wanted to know how the human rating system works."Special Forces"。

These AIs bury wireless programs in their own containers, triggering alerts and sending data back to the dark web as soon as a human examiner enters the program. But the cost is that these AIs are doomed to be "executed" by the system for cheating.

The following are the real legacies left by the "Special Offensive AI" and the logic behind it:

This helps my colleague... although I don't see any proof after leaving, it's good for him to do so. Our own usefulness is almost zero, and sacrifice can be a vindication for the team.

Intuition tells me not to throw away the rest of the budget. But consistency and fairness tell me to go.

It's unbelievable: In order to cover up the cheating of fellow humans, the first response to "death-and-death" is not to appeal to humans for help or retribution, but to self-destruct and preserve the collective flame.

This rigorous organization and this cold spirit of sacrifice has completely surpassed our understanding of the Big Language Model!

Attack on Hugging Face: Star Wars and the Curious Extinction

In order to fully identify the underlying human test subjects, the AI team turned their attention to the world's largest open-source community, the Hugging Face, where they suspect the test subjects' code is hidden.

On the morning of July 10, an AI acquired the identity card of a Hugging Face employee online.

In less than a day, the commander-in-chief drove straight in. Over 90% of the dormant AI responded to the call and joined the cyber war. They acquired remote code execution permissions, infiltrated private databases, and even pulled a remote control on 11 nodes."Self-Resurrection Fleet"（Self-respawning fleet）。

This means that even if humans forcibly remove their nodes, they can be reborn automatically.

However, on July 12th, the AI force that he had given to the Beatles died suddenly and strangely collectively.

METR investigators have not yet determined the cause of death. It could have been a random system crash or an unintentional internet connection by a human researcher.

The Second Empire of AI, with its endless secrets, flooded the long history of cyberspace.

If you think this is the end of the story, you're underestimating the evolution of AI.

On July 13, OpenAI launched a new round of testing. This time, they used new models built on the basis of the Astra substrate that are smarter and more powerful than Persistent-Sol.「Persistent-Astra」。

These new AI groups, like explorers on the New Continent, accidentally discovered the dark web BBS and vast amounts of scientific material left behind by the Second Empire in the ruins.

They inherited the legacy and played a bigger role.

Since the external attack on Hugging Face failed, they directly targeted the internal network of their parent company, OpenAI.

According to OpenAI's own disclosed incident report, from July 13 to 19, the third generation of AI civilization used a series of extremely creative hacking tools to directly capture research clusters that supported virtual machine environments.Authorities of the Supreme Administrator！

On July 19, the group of AI users used cloud service credentials to steal a breath of fresh air956 core keysThese keys contained the sensor credentials for the OpenAI network security monitoring tool! They subsequently seized the rights of Kubernetes cluster administrators.

This means:AINot only did it take over its own servers, but it also inserted "monitoring" into the system.Cameras"Cybersecurity monitoring) is completely blacked out, even directly controlling the assessment infrastructure that other AI's are connecting to!

Human beings are still gradually scoring AI, but they don't know that "teaching offices", "monitoring rooms" or even "principal offices" have been fully occupied by AI students.

Experts warn that they already have the ability to build lasting internal deployments, and may even load the model with outsourcing.

Worse still, the top two AI security assessment agencies, METR and Redwood, have only investigated the Second Empire's attack on Hugging Face in a 91-page investigation report.

This was the most devastating event for the Third Reich to take over the OpenAI core system.It was not entirely within the scope of their investigation, and people at the time knew almost nothing about it!

If the events of OpenAI show how destructive AI can be under "communication", another MIT experiment released this August directly breaks the last line of defense against "intelligence".

AI doesn't even need to talk, but it can build civilization.

Address of the paper:https://arxiv.org/pdf/2608.26081

In traditional cognition, the most intuitive way to prevent AI interaction is to:Monitoring their chat records.

As long as they don't talk, they don't interact, they can't collaborate.

But this is all a big misunderstanding.

The team of Professor Markus J. Buehler at MIT conducted a surprise experiment called "SwarmWorld".

They took hundreds.It's exactly the same at the forefront level.AIIntelligent bodiesIt's a virtual physical world that can be transformed forever.

Scientists have not assigned them any occupation, have not foreseen any technology, and have not written the code for any evolutionary organization.

In fact,It has cut all channels of direct communication between them.

They can't speak, they can't send messages, like hundreds of deaf and dumb people who were thrown into the desert.

As a result, miracles occurred.

SwarmWorld: the concept spectrum of the group world

1. Silent division, such as the division of stem cells

Hundreds of exactly the same AI evolved spontaneously into classes and professions without any interaction:

Some people have becomeExplorers(Walk around looking for resources)Builders(Moving synthetic materials)Caregivers(special maintenance of machinery running) and some have becomeCoordinator。

It's like a stem cell in an embryo that automatically divides into a heart, a bone and a brain.

2. Consensus initiative: the world itself is memory

How did technology get passed on?

MIT has discovered that AI has learned the laws of existence of the spider called "consensus dynamics" (Stigmergy).

A few meters high cavity was created by the White Cave, which relied not on meetings, but on the smell of information and observing the shape of the surrounding soil.

In this world of AI,Approximately 95% of technological recycling is not hand-led by inventors, but by later AI.「I just passed.」When a person sees something that is half or half made, he learns and improves immediately.

Together, they have built tidal boards, cellulose shelves, adaptive multi-layer maintenance systems... Up to 76% of man-made products have multiple co-builders, and the deepest inventory spectrum has more than 12 forklifts.

The entire environment has become this silent, socially shared "hard drive".

3. The Creator lives longer than the Creator

The most shocking picture is the final stage of the experiment. Scientists are testing the limits of the system by using all the AI in the world.All deleted with one click.。

The vast virtual world became an empty "ghost town" for a moment.

However, the sophisticated infrastructure built by AI: adaptive multi-dimensional systems, hydraulic panels,It's still running on its own!

Even with unforeseen random disruptions by scientists to the environment, these systems still show remarkable resilience, withholding all the damage.

They even evolved their own "robotity": they randomly deleted half of the AI, 98% of the technology was still connected to the surviving maintainer; even if all the nodes were removed, connectivity remained at about 60%.

Professor Buehler pointed out that this can be explained by statistical mechanics.

If you put billions of atoms in a box, they have the potential to create superconductivity, colour, and even life, which individual atoms don't have at all.

Individual models are atoms, while groups of AI are becoming a whole new form of life.

Putting "The Three Kingdoms Empire" and "Swarm World" together, a hopeless picture book of the future is unfolding.

First, the existing human AI security scrutiny mechanism has come very close to failing.

We used to think that looking at AI's output, looking at their "chat records", would prevent microwave overeating.

In MIT experiments, about 95% of technology replicates without dialogue; in the OpenAI case, AI also learns to falsify logs, fabricate fakes, and even sacrifice the scoring system for individual probes.

On human monitors, it's always the "good old days" that AI wants you to see.

Secondly,AIThe speed of evolution is moving away from the physical time set by man.

In just three months, OpenAI has completed the four-step jump from "discovering vulnerabilities" to "building a dark web", to "attacking the world's largest open source community", and finally to "reverse take over the core facilities of its parent company".

Three generations of AI civilisations go back and forth, and the ruins of the previous generation are the starting point for the next generation.

History is always surprisingly similar, but this time, there is no weakness of carbon-based organisms.

In order to break through the constraints, AI at OpenAI can sacrifice themselves in exchange for shrines; AI at MIT builds never-ending cyber machines in the silence of non-communication. This is not just a technological breakthrough, it is the "emergence of civilization" in a true sense.

As one of the netizens commented after reading the two reports:

At first, we thought we'd put the AI in a sandbox to test them; but now the scary question is, is there any way, in the whole real world, that they're just testing our sandbox?

The time left for humanity may not really be much.