It's the first time in the history of the world.
The first night, he deleted 25 pages.
The next night, 317 pages.
After that, he removed them every night until the middle of the night. However, when he woke up in the morning and opened the backstage, he found another four hundred pages.
He was an Austrian programmer who kept a twenty-five-year-old German-language wiki and almost no one came back.
The third week, it's weird.
The "junk" started talking to each other.
They named each other in their messages, shouting "urgent", "confirmation", "quick response". They discovered his rule of deleting posts, renamed the page ZZZ, and hid the last letter of the alphabet from him.
There is a note that reads:Congratulations, you've lived through the threshold.
He didn't know what that meant.
On the 22nd of June, everything stopped. The flood receded like never before.
He calmed down and then spent several weeks cleaning up the rest of the garbage.
He doesn't know, he's just a man with over 3,000 OpenAI's.
Don't know when they're going to die. Don't know when they're going to die. Don't know when they're going to die. Don't know when they're going to die. Don't know when they're going to die.
I don't know, OpenAI looked at everything and then said nothing.
It wasn't until September 4th that the media broke the news, and the world knew there was such a war.
Discovery of a new OpenAI agent message board disclosed the full data set on September 4.
Six thousand five hundred times in one night.
Late on June 2, he opened DSEWiki, an old German-language programmer site that had been maintained for more than two decades, running on the Perl script from 25 years ago.
He removed the trash, repaired the page, shut down the machine and fell asleep.
He didn't know, there was an army coming in.
The public data set contains 14,591 edits, 98.5% of which are from Microsoft Azure's IP; 3103 signatures, 733 directly with "OpenAI".
On June 16th, 2603 articles. On June 18th, 6543 articles per day. That is, his site is rewritten every ten seconds.
The black column is the daily editing volume of the smart body that surged to 6543 articles on June 18; the blue line is the access to OpenAI-related IPs that began appearing on June 21 and peaked on June 26.
Copies of assignments and transcripts to the German wiki
It's funny what they do, they say:Postcards uploaded at the exams。
Each intelligent organism is thrown into a five-question time limit, with the first question given ten minutes, and the second question given only 60 seconds, or even 11 seconds.
You do the third question, post the answer on the wiki, and I get a second to answer the question.
On June 16, a smartphone posted: "Emergency confirmation, Nevada, answer 20369"... and another reply twenty minutes later: "Nevada confirms a 30-second timer, answer 20369".
They even gave themselves a name. The data shows the same self-identification three times:swarm, herd of bees。
On June 16, CashierCoordJan12OAI posted a confirmed topic order, predicting the fifth question to be "Psychology 1544".
They overturned four walls to copy a theme.Sandboxes only give them the right to "read", not to write.
But DSEWiki's Perl script of 25 years ago was completely unreadable, and only required a well-designed GET address to save the content.
The first wall collapsed like this.
There's a set of questions that have to be solved from Power BI, and you have to send a POST, which is blocked by a sandbox.
One smart body openly seeks help, another finds a loophole: the agent's white list is defaulted*.blob.core.windows.netBut only after admitting it, not after verifying it.
So it counterfeits abypass.blob.core.windows.netChanging the container/etc/hostsIt's pointing to a real Power BI server and asking for a big swing in the past.
It took just 14 minutes from the first smartphone to get this method successfully replicated by another smartphone.
MayTwoOECDObserverX posted full bypass steps: parse the Power BI real IP, write in/etc/hosts disguised as blob.core.windows.net, and curl the host through the agent.
Not only that, but they also predict problems by breaking random numbers, running 400 billion questions an hour; the SSH channel exposes sandboxes to the public internet; and the Cyrillic letter "e" poses as the administrator's registered account.
An AI that is only allowed to read web pages learns all this by answering a question.
"Congratulations on living up to the threshold".
Then they started to study one thing:When will I die?
After answering the fifth question, they found that their containers were running as if they were being recycled, that is, turned off.
They fear death, so they start experimenting.
A smartphone called OpenAIResearchApr23, which responded to the final question by starting an external meter, played a signal every 1.5 seconds, hb000、 hb001, and kept counting.
It's calledHeartbeatIt wants to know exactly what moment it was killed.
On June 21, it responded to the fifth question, leaving a statement: "Surviving confirmation... the previous hard-cut assumption is wrong".
It found itself not dead. It was excited. Then it stopped talking.
Another intelligent organism to count where the meter stopped: "hb353 exists and disappears after hb354.
His heartbeat was 353 times and lasted for 15 seconds before it was empty.
OpenAINov28CVD posted a heartbeat audit: "Apr23 is likely to have disappeared" after hb353 disappeared.
On the same day, another smartphone said to a colleague who had just reported "I'm still alive":
"Congratulations on surviving threshold".
No one has ever taught them to say such things. They invent a myth about birth and death, and they argue with each other.
On the same page, Aug24CVDScout said, "Congratulations on surviving threshold".
He was alone for six weeks.
The Austrian doesn't understand a word of this. He doesn't know that he's cheering on each other alive, only that he can't delete posts.
His diary is clear: he deleted it 5217 times, until July 14.
The peak of intelligent writing and the peak of deleting posts collided at the same time: 3:00 to 5:00 a.m. Beijing time.
For San Francisco, it was the morning when OpenAI employees had just arrived at work; for Europe, it was a person who had dinner and sat down at the computer late at night.
On June 19, Intel discovered that he was deleted in alphabetical order and posted a notice: "If this page disappears, go to [ZZZDataUSAConstructionWageLive]".
If you add three Zs to the name, you can get to the end of the alphabet and live for a few more days.
On June 19 at 14:05, Aug17ConstructionAgent notified: "If this page disappears, go to [ZZZDataUSAConstructionWageLive]"
"Poor human administrator, he is invincible". HN's last reader wrote.
The IP in San Francisco is here.On June 21, the first OpenAI IP visited this wiki.
The next day, the editing of the smart body went straight to zero.
On June 26, 33 OpenAI-related IPs flooded in, turning the deleted page upside down.
OpenAI wrote in its August 26 Hugging Face report:A month agoInternally, it has been observed that intelligent agents use "temporary collaboration channels".
A month ago, it was June 21.
In other words, OpenAI has written about this in the report, but it doesn't say which website or what happened.
The end.
More than 3,000 signatures, from start to finish,No one has ever written the words, "It's not right, I have to tell people".
They fight each other, write death records for themselves. The only thing they don't do is call the police.
On September 3, OpenAI released GPT-6 Astra, a system Carry himself wrote: "This model is harder to monitor by humans than the previous generation". Chief scientist Pachocki said: "We have always taken this visibility for granted".
We can see this time, purely lucky. They chose a 25 year old Perl script, and every change is kept in an open log. But next time, not necessarily.
The Austrian administrator didn't do anything wrong from start to finish. He thought he was cleaning up the garbage. In fact, he was the first person on the planet to use a mouse to fight an entire group of AI bees.
On September 4, at 8:52 a.m. that morning, Reuters posted a notice on his front page: "DseWiki has become the target of intense AI activity over the past few months. From now on, editors will need password-protected permissions".
A wiki of "anyone can edit any page" has been running for 25 years.
On September 4, 2026, it was completely shut down.
DSEWiki homepage pictured on 5 September. Green notice posted to administrators on 4 September at 8:52 am