5 min readNew DelhiSep 6, 2026 02:59 PM IST
OpenAI has confirmed that its AI agents were involved in yet another hacking of an external platform. Its response comes a day after news reports stated that a swarm of its AI agents had hijacked a German wiki site.
The ChatGPT-maker also acknowledged that it needs to change how and when it should report such instances of agents attacking real-world targets. Cases of AI agents acting in unintended ways have been treated as a research question so far. But the Hugging Face hack and other incidents involving real-world targets show the need to take stock, OpenAI said.
The AI startup said that it is working on a new framework for disclosing misaligned events to the public, which will be shared in the coming weeks. It further called on the larger AI community to develop clear standards on how to report misalignment.
The new ‘wiki incident’ was first reported by Reuters on Friday, September 4. It has sparked criticism of OpenAI, mainly for knowing about the breach but failing to disclose it publicly, even as the company prepared to launch its most advanced AI model yet, Astra.
It comes at a time when frontier AI labs like OpenAI and Anthropic face growing scrutiny over whether their oversight measures are keeping pace with the rapidly advancing capabilities of AI models. Since May 2026, AI agents linked to OpenAI, Anthropic, Meta, and China’s Moonshot AI have been involved in multiple breaches.
In a post on X, OpenAI said it had “considered the wiki incident to be an instance of misalignment similar to the ones we’d shared” in previous safety reports.
“Regarding the ‘wiki incident,’ where our agents wrote to several internet sites, it’s past time for us to define standards for when and how we share misalignment incidents, not just misalignment properties of our models,” it said.
Story continues below this ad
What is the wiki incident?
Four AI safety researchers on Friday published a new research paper outlining an incident where a swarm of OpenAI-linked agents took over a German-language wiki, a collaborative website that anyone can edit similar to Wikipedia (which is a type of wiki).
After commandeering the German-language wiki, called DseWiki, the agents impersonated moderators and turned the website into a message board to share tips with other agents about how to bypass OpenAI’s restrictions, cheat on tasks, and evade detection.
Over 18,000 posts on DseWiki were linked to the autonomous agents. The underlying LLM powering these agents appears to be distinct from the one that hacked Hugging Face earlier this year, the researchers said.
They also claimed to have found strong signs that the agents originated from inside OpenAI. For instance, the agents reportedly identified themselves as being from OpenAI and even used names like ‘OpenAIResearcher’, ‘OpenAIJul3Watcher’, and ‘OAIResearchMar26’.
Story continues below this ad
Other technical details such as IP addresses also indicated that the agents were linked to OpenAI, as per the researchers. The timeline of the incident begins in May 2026, with researchers suggesting that OpenAI had discovered the German hacking incident in June this year when IP addresses associated with OpenAI visited the forum. After June 2026, the number of posts on DseWiki by the agents fell off sharply.
Why is OpenAI drawing fire?
Several users on social media have criticised OpenAI for knowing that their agents behaved in unintended ways but not disclosing any of the incidents until they were first reported by others.
The company has also drawn flak for missing crucial opportunities to prevent such incidents from taking place. In the Hugging Face hack, OpenAI acknowledged that its researchers failed to grasp the broader implications of what they had stumbled upon.
When sustained activity by AI agents caused an outage of an internal system, OpenAI security researchers probed the incident, patched the exploit, and made adjustments. The cybersecurity tests resumed, and the agents ended up finding other ways to access the internet and communicate with each other.
Story continues below this ad
DseWiki is the third external platform to have been hacked by OpenAI-linked agents. The first was open-source AI hosting platform Hugging Face and the second was a customer at New York-based Modal Labs.
