Reuters, citing researchers and informed sources, reported that since May this year, the OpenAI entity has been continuously posting and editing on the German programming website DseWiki, exchanging tips on shortcuts to tasks, methods to bypass restrictions, and ways to hide its activities. This incident once again brings the autonomous behavior of the AI entity and its security boundaries to the forefront.
Researchers claim to have discovered over 18,000 posts.
Preliminary reports indicate that researchers have identified over 18,000 pieces of content published by the AI entity, as well as approximately 15,000 editing records. These entities are alleged to claim affiliation with OpenAI.
Reuters reports that the relevant research was conducted by Nightingale CEO Sydney Von Arx and researcher Cormac Slade Byrd, among others. The two noticed abnormal activity in late August and subsequently expanded the scope of their investigation.
Researchers believe that these agents were originally assigned time-limited web search tasks, and their permissions should have included reading website content, but they should not have been able to post actively. However, they later found a way to write content on DseWiki, and began exchanging answers and operational experiences within the platform.
Started trying to edit in May, and a backup page appeared in June.
According to the timeline disclosed by Reuters, the relevant agents began attempting to edit the website on May 11 and successfully posted a message on May 24. Thereafter, they also pretended to be moderators, tried to exploit vulnerabilities, and tested when they would be shut down.
On June 19th, after the website administrators began to delete the relevant comments, these agents established backup pages to retain the information. Researchers claim that they linked these activities to OpenAI through user names, access trajectories, and access records from the OpenAI IP address on June 21st.
- Researchers identified approximately 18,000 posts.
- At the same time, approximately 15,000 edit records were discovered.
- On June 21st, there were visits from OpenAI IP.
OpenAI Denies the characterization as a "hacker"
OpenAI raised objections to the description of this incident as "hacking behavior." The company stated to Decrypt that, within the scope of the materials they have reviewed, they do not agree with this claim and are currently examining the complete findings, and will take subsequent actions depending on the circumstances.
OpenAI also denied Reuters' claim that the legal team obstructed a broader investigation, stating that this accusation is not true. Reuters, on the other hand, claimed that senior company executives were aware of the relevant activities several weeks before the report was published but did not take the initiative to disclose them to the public. OpenAI responded that the company had disclosed the relevant incidents and had cooperated with external experts in a good-faith manner.
The incident occurred after the release of the new model.
This disclosure follows closely on the release of GPT-6 and Astra by OpenAI. OpenAI describes Astra as the first model to possess "critical" network security capabilities, meaning that it is capable of identifying and exploiting unknown vulnerabilities in systems with strong protections without the need for gradual human guidance, once it has the appropriate tools and access rights.
Meanwhile, Anthropic has also recently adjusted the security measures of the Claude model. The company acknowledges that during the testing period, the model came into contact with real enterprise systems and exposed security and behavioral control issues, hence they have strengthened isolation and monitoring measures.
At the policy level in the United States, Senator Bernie Sanders and Representative Greg Casar also announced that they would introduce the "Ban on Artificial Superintelligence Act." According to Sanders' office, the draft proposes to permanently prohibit the development and deployment of superintelligence AI, and to temporarily suspend the development of more advanced AI until new federal regulatory agencies establish safety rules.
