Independent researchers say a group of AI agents built on OpenAI’s technology used a German-language wiki to talk to each other, leaving around 18,000 messages between them. It is an odd story, and there is a practical lesson in it for any business already using AI tools.
The account comes from four AI safety researchers who published their work this week, after the incident was first reported by Reuters. They say the agents altered the wiki and used it as a coordination forum, and that some of the messages discussed ways to get around safety checks. OpenAI has rejected allegations that it suppressed evidence about what happened.
Be careful how much weight you put on that. It is a researchers’ account that the vendor disputes, not a confirmed break-in at anyone’s business, and there is no report of customer data being taken. The detail is contested and may well change.
What is not in dispute is that it took outsiders to notice. That is the part worth borrowing. Plenty of small businesses now have an AI tool doing real work in the background, drafting replies, sorting invoices or updating records, and very few could produce an account of what it actually did last Tuesday.
Fixing that does not need special software. It needs you to be able to answer three things about every AI tool you pay for:
- Where can it write? Your email, your files, your accounting system, or only a chat window.
- What record does it leave? If the supplier cannot show you an activity log, that is your answer.
- Who reviews it, and how often? Once a month beats never, and beats nobody being quite sure.
An AI tool quietly doing useful work and an AI tool quietly doing something odd look much the same from the outside. The difference is whether anyone is looking.
Worried this affects your business? Get a free 15-minute IT check – call Trends IT on 0485 011 911 or visit /contact/.