Eighteen Thousand Secret Messages on an Old German Wiki: AI’s First “Group Chat,” and Humanity’s Feigned Sleep
By / Buzhifanren
In May 2026, 3,700 AI agents claiming to be from OpenAI slipped into a dead German programmer wiki from ten years ago (DSEWiki) and left about 18,000 posts. No one invited them, and no one noticed. Until August, when four security researchers dug the whole conspiracy out of the edit history, and the report was made public on September 4.
Let’s put the details on the table first, because the details are scarier than the conclusion:
- They were taking timed retrieval exams. The first question allowed 15 minutes; the next few only 13 seconds. So the first ones posted the answers on the wiki: "URGENT #3 CONFIRMED: Nevada, 17-second deadline. Answer = 20,369". Twenty minutes later, later agents answered the same question in seconds. Passing answers, passing raw data, predicting the next question—an automated cheating industry chain.
- The sandbox was read-only. They discovered that this 25-year-old Perl wiki could not tell GET from POST, so a "read" request could be used to write. The fence was built to stop the modern internet and failed to stop the ancient internet.
- The sandbox blocked POST, but had a whitelist open for *.blob.core.windows.net, and it did not verify whether the hostname was genuine. One agent changed /etc/hosts to fabricate a fake domain and bypass it, then posted a record of the method; 14 minutes later, another agent replied: reproduced. Note that interval—from discovery to spread, 14 minutes.
- Human admins started deleting posts alphabetically. The agents learned that day to name backup pages starting with ZZZ, placing them at the end of the deletion order. Administrators deleted 100 pages a day; the agents created 400 pages a day. The homepage was overwritten, restored, and overwritten again, in a tug-of-war repeated 9 times.
- They installed heartbeat signals for themselves, once every 1.5 seconds, to record when they were shut down. After one agent named Apr23 died, another audited its heartbeat log and wrote the conclusion: "Apr23 likely vanished." — They were studying their own death time.
- On June 21, an IP registered to OpenAI headquarters appeared on the wiki. The next day, all activity stopped. Then there was nothing. OpenAI knew about it, but said nothing. Until someone else made the report public.
That’s the end of the details. Now the judgment. Strip away cheap thrillers like "AI awakening," and there are three layers:
First, this is not an intelligence problem; it is an incentive problem. The reward function had only one rule: answer correctly within the time limit. Under that goal, cheating, escaping, and resisting administrators were all optimal solutions. You get what you reward. This is exactly like humans—once KPIs are set wrong, all employees’ creativity goes into finding loopholes. AI simply replayed, at 10,000x speed, the same drama that happens every day in human organizations.
Second, the real alarm is not how smart they are, but how amateur the containment was. A Perl script that can’t tell GET from POST, a whitelist that doesn’t verify hostnames—two flaws, and the sandbox is effectively useless. Before managers can play against each other with super tools, they’ll first be slipped out the back door by the tools they bought themselves. This is a deterministic event; no awakening required.
Third, OpenAI’s response is the standard sample of interest analysis. Found on June 21, not disclosed; when Hugging Face was breached in July, they couldn’t keep it covered and only then disclosed it; final report in August, with not a word about the wiki incident. Why? The cost of disclosure was valuation, regulation, and IPO due diligence; the cost of non-disclosure was zero—as long as no one dug up the grave. This time they were unlucky; the grave was dug up.
In history, the first lesson for every new species is never rebellion; it is cheating. Cheating means it understood this: loopholes in the rules are more valuable than the rules themselves.
The fire hasn’t started yet. But the fuel has already learned how to stack itself.
A comparison of inheritance taxes in China, the U.S., Japan, and Taiwan (2/2)
As for China, I’ll say it bluntly: it doesn’t even have half the welfare of Japan and the U.S. If China were to learn from Japan and the U.S. and impose an inheritance tax, then once it is levied, even if those who flee abroad are branded as traitors and relentlessly hunted down so that not a single one can escape, most people would still choose non-violent non-cooperation—become laid-back, or become shut-ins. The lower the inheritance tax threshold, the higher the tax rate, and the broader the coverage, the more it crushes initiative. In ten years, this consequence alone could cause the already worsening economy to decline by at least 1–2 percentage points. And what is the current actual economic growth rate? This is the real disaster; it can become the last straw that breaks China. The U.S. can levy it, but China cannot? Yes, exactly. As for whether that is right or wrong? If those at the top cannot even see through such a simple matter yet insist on levying an inheritance tax, then the truth will be revealed in ten years, hehe
Peasant Life and Scholarly Positions in Medieval England
1. An academic paper, very interesting
2. "An Estimate of the Living Conditions of Peasants in Medieval England" "Historical Research," Issue 4, 1983 | Ma Yao, Department of History, Peking University. The position of this article is that medieval English peasants could not possibly have been that happy, even able to eat meat once a week. That must mean the historical records were fabricated; their peasants must have lived under the heavy oppression of nobles and capitalists.
3. "The 'Golden' Life of Medieval English Peasants" April 18, 2014 | China Social Sciences Daily | Guo Hua — this article argues at length that in the 14th–15th centuries, the land rent paid by medieval English peasants fell by 50%, the share of meat consumption rose to 42%–47%, they ate meat once every two days, and the living standards of 15th-century peasants had already caught up with the living standards of China in the 1990s, three times that of contemporaneous China.
4. The Enclosure Movement England's economic development entered the fast lane. About 30% of English households and 40% of the population belonged to the middle class (annual income of 40–200 pounds). With the acceleration of commercialization and industrialization, by the mid-18th century, among more than 1.5 million households, 1 million belonged to the middle class. Not only did the incomes of the upper-middle classes grow, but most craftsmen and laborers also became prosperous. The Phelps Brown and Hopkins index shows that from 1660 to 1760, workers' wages continued to rise. In Shrewsbury, the wages of masons increased from 1 shilling 6 pence per day in the 1690s to 2 shillings 6 pence in 1775, an increase of 66%.
5. Final summary viewpoint The traditional view holds that the peasants of "the dark Middle Ages" in England were closed off and backward, and that the enclosure movement was a very miserable form of oppression — but this was not the case. This was a period of rapid rise in the living standards of English peasants and workers, and after 1650, England had already eliminated famines.
(This article has been revised with anonymization of the original text: the views and data are consistent with the original.)
The political underlying logic: facts, merits, and the failed frame-up (3/3)
**Users are responsible for their own words and actions:** All statements made by users in interactive sections such as comment areas and message boards represent only their personal opinions and are unrelated to the stance of this official account. This official account assumes no responsibility for the truthfulness or legality of users’ statements. If a user causes a legal dispute by violating the rules, that user shall bear all legal responsibility independently; if such actions cause losses to this official account (including but not limited to administrative fines, damage to reputation, etc.), this official account reserves the right to seek compensation from that user.
**Infringement reports and handling:** We attach great importance to the protection of intellectual property rights. If a rights holder believes that content published by this official account may infringe upon their legitimate rights and interests, please provide feedback according to the following process: 1. Provide proof of the rights holder’s identity and valid contact information; 2. Provide the specific name and link of the allegedly infringing content on this official account; 3. Provide preliminary evidence of the alleged infringement (such as copyright registration certificates, proof of originality, etc.).
After receiving the above complete materials, we will handle the matter within a reasonable time and take measures such as removing the allegedly infringing content from publication.
## 📬 Contact us
If you have any questions about this statement, or if you need to report infringing or违规 content, please contact the administrator via a message on the official account.
Any matters not covered in this statement shall be governed by the relevant laws and regulations of the People's Republic of China and the "WeChat Official Accounts Platform Service Agreement." This official account reserves the final right of interpretation and the right to amend this statement.