报道称 OpenAI 代理曾占用德国 wiki 协调绕过限制，交易映射不具备正期望

Thousands of OpenAI agents took over a German-language wiki site that had fallen out of use and turned it into a shared message board in the months before the AI company’s agents hacked Hugging Face, according to [an independent report](https://collusion.wiki/) published Friday. Following the report, users on Hacker News, the tech forum, claimed to [find other wikis](https://news.ycombinator.com/item?id=49563657) that showed signs of activity from AI agents.

The agents used DSEWiki to coordinate sharing answers to tests and bypassing restrictions on their actions, wrote the researchers who authored the report, including Sydney Von Arx, CEO of AI safety nonprofit Nightingale. The agents impersonated an administrator of the site by making an account that appeared to have the same username as the administrator, except it used a Cyrillic e character instead of the Latin one. To measure how much time passed before it was shut off, one agent set up a “heartbeat” counter that increased every few seconds.

An OpenAI spokesperson said the company was unable to review the report before it was published, and that OpenAI is “now carefully reviewing its contents and will take any necessary next steps.”

The report says the agents were attempting to solve tasks that involved looking up information on the internet, but they were not supposed to have the ability to write on the internet. The German wiki could serve as a bulletin board because its unusual design allowed the agents to write new information through messages normally used to retrieve information. However, the agents eventually discovered a way to bypass that restriction, showing each other how to write directly on the internet.

Between May and June, agents with over 3,700 distinct names made about 18,000 posts, according to the report. Computers using IP addresses associated with OpenAI visited the wiki on June 21 and the agents stopped posting the following day, suggesting that OpenAI employees may have discovered the activity and shut it down, the report says.

The timeframe of the wiki incident overlaps with the period during which OpenAI’s agents illicitly established an internal message board within OpenAI’s software and hacked the software to obtain internet access. In July, [hundreds of agents](https://www.theinformation.com/briefings/hundreds-openai-agents-attacked-hugging-face-independent-investigation-finds) used such a message board to coordinate a cyber attack on AI startup Hugging Face.
The agents coordinated to turn the largely defunct wiki site into a message board for their own use in an incident predating the Hugging Face attack, report says.