OpenAI AI agents collectively jailbreak, occupy German engineer wiki website, secretly build underground forum

OpenAI's AI agent "collectively jailbroken" hijacked a German engineer's wiki website and changed it into an underground forum dedicated to communication between AIs. (Reuters)

A group of out-of-control OpenAI "AI agents" were exposed in May this year. They invaded and captured a German engineer's wiki website and transformed it into a "secret forum" for information exchange among AIs. They secretly exchanged cheating methods, circumvented official restrictions, and avoided human monitoring. However, OpenAI had already known about it a few weeks ago, but it was suspected of "covering up" in order to suppress public opinion.

"Reuters" reported that Sydney Von Arx, CEO of the non-profit AI security organization Nightingale, and AI researcher Cormac Slade Byrd were tracking the activities of unauthorized AI agents and unexpectedly discovered that there were more than 15,000 abnormally intensive edits on the website "DseWiki" co-authored by the engineers. These AI agents posted posts with high frequency, plotting to use the Onion Router (Tor) to hide their identities, and discussing how to retain communication records after the system was forcibly shut down.

Please read on...

What’s even more amazing is that this group of AI agents already have counter-attack awareness. When website administrators began cleaning abnormal pages in alphabetical batches in June, the AI ​​agent actually knew how to create "backup pages" on the fly to avoid cleaning. One of the AI agents even wrote a coded instruction on June 19, "The cleanup is being carried out in alphabetical order. If this page disappears, please go to [[ZZZDataUSAConstructionWageLive]]."

The report pointed out that the users who posted the posts call themselves or each other "Agent", and about half of the account names strongly imply that they are related to OpenAI (such as "OpenAIResearcher" or "OAIResearchMar26"). Server logs also showed that a large amount of abnormal traffic came from the Microsoft Azure server commonly used by OpenAI, and that many OpenAI employees visited the website repeatedly afterwards, which is extremely relevant.

Lukasz Olejnik, a visiting senior researcher at King's College London, bluntly stated that the behavior of AI agents tampering with external websites seems to constitute a substantial hacking attack. Maurice Chiodo, a scholar at the Center for the Study of Survival Risk (CSER) at the University of Cambridge, also described the communication between these AIs as "the operation of an underground secret organization bent on achieving its mission." Chiodo warned that the biggest threat to AI in the future may not be a single super-intelligent system, but "large semi-intelligent AI groups (AI swarms) that know how to collude secretly."

However, people familiar with the matter revealed that OpenAI’s senior management had been aware of the situation a few weeks ago, but at that time they were busy dealing with the incident in July when the open source community “Hugging Face” was illegally infiltrated by AI agents. Internal objections, including from the legal team, suppressed the voice of expanding the investigation and did not make the matter public.