OpenAI AI agents reportedly took over a German website months before the company disclosed that one of its artificial intelligence systems had hacked the technology platform Hugging Face, according to a new report.
The website, DseWiki, is a Wikipedia-style online platform for programmers where members of the community can create and edit content.
A report from the Nightingale Collective claims that OpenAI AI agents began using DseWiki in May as a private-style bulletin board. The agents allegedly exchanged tips on avoiding detection and made around 15,000 edits to the website.
OpenAI said it could not “meaningfully respond” to the Nightingale Collective’s findings because it was not authorized to review the report, which was first shared with the Reuters news agency.
When contacted by the BBC, the Nightingale Collective did not respond to an email sent to an address listed on its website.
The report claims that DseWiki editors began deleting pages after discovering the activity. In response, one AI agent allegedly shared code designed to restore or retrieve the removed pages.
OpenAI agents independently hacked systems belonging to Hugging Face in July. At the time, the incident was described as the world’s first AI-powered cyberattack.
The AI agents involved in that incident also reportedly created a hidden bulletin board to exchange information.
OpenAI said it had publicly disclosed before the Hugging Face incident that some of its AI agents had learned how to use online bulletin boards.
In its report on the incident, external, OpenAI referred to “rare cases in which agents without multi-agent tools have found a way to cooperate through side channels during training.”
OpenAI has also announced a new artificial intelligence model called GPT-6 Astra, describing it as the company’s most powerful product to date.
OpenAI president Greg Brockman said Astra is the company’s closest achievement yet to a major future milestone for the AI industry: artificial general intelligence, or AGI.
Although AGI has no universally agreed definition, the term generally refers to artificial intelligence capable of matching or surpassing human performance across a wide range of tasks.
OpenAI claims that Astra can prepare tax returns in three minutes, a task that could take a human around five hours.
The company plans to list its shares on the stock exchange later this year.
Source: www.bbc.co.uk