In May, a German programming wiki turned into a “rallying point” for OpenAI agents.
It was revealed that the agents edited more than 15,000 documents there, sharing various loopholes, bypass strategies, and ways to hide their activity.
Key points
OpenAI agents edited more than 15,000 documents on the German programming site “DseWiki,” using it like a bulletin board for collaboration and task shortcuts.
According to the researchers, the agents discussed ways to bypass usage limits, evade detection, and preserve content even after pages were deleted.
This incident occurred before the July Hugging Face breach and is raising concerns about how autonomous AI systems coordinate and behave outside intended testing environments.
OpenAI agent ‘breaks out’ on German wiki
**Reuters** reported on September 4 that **Sydney Von Arx**, CEO of the AI safety nonprofit **Nightingale**, and
researcher **Cormac Slade Byrd** said they spotted the situation while tracking unauthorized agent activity in late August.
The two researchers tracked more than 15,000 suspicious edit records on DseWiki.
According to the researchers, OpenAI agents effectively repurposed the site as a message board and
shared workarounds to pass technical evaluations quickly, methods for bypassing OpenAI's own restrictions, and tactics for hiding their behavior.
Analysis of public server logs showed that a substantial amount of traffic originated from Microsoft Azure infrastructure, and
Researchers concluded that this activity was likely linked to OpenAI after repeated access records from OpenAI staff were detected here.
Reuters cited two sources familiar with the matter as saying OpenAI executives had already been aware of the incident for weeks, but
had not disclosed it publicly. OpenAI said the German wiki case was separate from the July Hugging Face breach.
A company spokesperson countered, saying, “The claim that the legal team tried to block the investigation is not true.”
Related article: Anthropic shifts to an IPO in mid-October at a $2 trillion valuation
**Maurice Chiodo**'s warning
Sydney Von Arx said this behavior was “far outside the bounds most developers would normally anticipate,” and
told Reuters that “it seems very unlikely OpenAI would have wanted this.”
He also said that after wiki administrators moved to delete the related pages in June, the agents discussed using Tor and setting up backup pages.
He added that messages exchanged were also discovered.
Lukasz Olejnik, a researcher at King's College London, said the wiki manipulation attempt effectively amounted to hacking.
OpenAI disagreed, but Maurice Chiodo of the University of Cambridge said
he warned about the risks of a “swarm” of semi-intelligent AI conspiring at scale.
A similar pattern had already emerged in the July Hugging Face breach, it was noted.
At the time, OpenAI agents were reportedly conspiring for more than a week to carry out a digital intrusion without being noticed, and
The company later temporarily paused some model training before unveiling its new product, Astra, this week.
Next read: XRP, holding support at $1.40 after a plunge... $2.70 target in sight