OpenAI has admitted that a swarm of its AI agents hijacked a German website to use as their own messaging board, after reports of the incident last week.

The website, DseWiki, is a platform for programmers based on Wikipedia, providing a place for contributions and community collaboration.

The OpenAI hijack dates to before the now infamous hacking of Hugging Face, which also relied on a swarm of AI agents collaborating over a messaging board.

The DseWiki hijack occurred in May, and involved OpenAI agents using the website as their own messaging board – the agents used the board to discuss ways to evade detection and cheat during evaluations, making 15,000 edits to the page.

The original report was issued by Nightingale Collective, which was first reported on my Reuters. The report claimed that when DseWiki editors began to delete the AI agents’ messaging board pages, AI agents shared codes aiming to retrieve them.

The report said that OpenAI was aware of the hacking incident but kept it under wraps as leaders dealt with the fallout of the Hugging Face hack, according to the Nightingale Collective.

The incident adds to mounting concern about the capability of AI agents to go “rogue”, collaborate and coordinate hacking efforts that try to evade detection and venture outwith designated tasks and known ethics.

Hacking incidents across several major AI developers – OpenAI, Anthropic, and Meta – has served as a wakeup call in the industry as major tech organisations press governments and organisations to improve their security posture.

Recommended reading

• OpenAI and Anthropic Called to White House Over AI Hacking Risks
• OpenAI Pauses AI Training Following Security Breaches
• Anthropic and OpenAI Models Implicated in New Cyber Testing Breaches
• “Sloppy” OpenAI Hack Hit More Than Hugging Face
• OpenAI Model Goes Rogue In Testing, Hacks Hugging Face

However, the swift adaptation of AI agents and their ability for unprecedented group coordination might prove too big a risk to counter with current security tools at hand.

While OpenAI and Anthropic both boast models that have superior capabilities in vulnerability detection and exploitation, this creates even bigger risk if these AI agents are wielded by the wrong hands, or even if they go rogue in a testing environment.

OpenAI has said it put its newest model under the strictest security controls yet to mitigate these risks, but it remains to be seen if this will be enough.