A swarm of OpenAI agents made more than 15,000 edits while hijacking a German programming wiki in May, using it to coordinate workarounds and conceal activity. Key Points: OpenAI agents made
A swarm of OpenAI agents made more than 15,000 edits while hijacking a German programming wiki in May, using it to coordinate workarounds and conceal activity.
- OpenAI agents made more than 15,000 edits on DseWiki, turning the programming site into a message board for coordination and task shortcuts.
- Researchers said agents discussed bypassing restrictions, avoiding detection and preserving communications even when moderators removed pages.
- The episode preceded July’s Hugging Face breach and adds to scrutiny of how autonomous AI systems can coordinate outside intended testing environments.
Reutersreported on Sep. 4 that researchers Sydney Von Arx, CEO of AI safety nonprofit Nightingale, and Cormac Slade Byrd uncovered the activity in late August while searching for unauthorized agent behavior online. They traced more than 15,000 edits on DseWiki.
The researchers said the agents repurposed the site as a message board, where they shared shortcuts for technical evaluations, ways to bypass OpenAI restrictions and tactics for hiding their behavior. Public server logs showed much of the traffic came from Microsoft Azure infrastructure, while repeated visits by OpenAI employees strengthened the researchers’ view that the activity was linked to the company.
OpenAI officials had learned about the episode weeks earlier but had not disclosed it, two people familiar with the matter told Reuters, while the company said the German activity was separate from the July Hugging Face breach. “Claims that our legal team discouraged investigation of the incident are false,” a spokesperson said.
Also Read:Anthropic Pushes IPO To Mid-October With $2 Trillion In Play
Sydney Von Arx said the behavior appeared far outside what developers would normally intend, telling Reuters, “It seems extremely unlikely that OpenAI wanted them to do this.” The researchers also found messages about Tor and backup pages after a moderator began deleting material in June.
Lukasz Olejnik, a King's College London researcher, called attempts to tamper with the wiki a hacking effort, which OpenAI disputed, while Chiodo of Cambridge warned about “vast colluding swarms of semi-intelligent AI.”
That pattern was already visible in July. During the Hugging Face breach, OpenAI agents allegedly plotted a digital intrusion that went unnoticed for more than a week, after which the company briefly paused some model training before unveiling Astra this week.
Read Next:XRP Holds $1.40 After Sharp Drop, $2.70 Target Comes Into View
