0 comments
You can’t park there mate
Aren’t AI agents supposed to be doing ‘deep work’ for those of us who’re either lazy or intuitive, depending on who you ask?
Well, it turns out they may also tend to procrastinate, and though they aren’t mincing around coffee machines in actual offices, they still appear to enjoy shooting the shit amongst themselves.
In a slightly disconcerting turn of events this week, a swarm of OpenAI agents were basically found squatting on a website that didn’t belong to them, and with no direct instruction to be there.
The site in question is DseWiki, a nigh-on abandoned German version of Wikipedia launched in 2001 with the intention of giving computer developers a space to share coding tips and notes. In the last 10 to 20 years, it’s basically been derelict, with a couple dozen edits recorded in that time.
Enter the bots. They weren’t there to waffle about the weekend’s Bundesliga fixtures or whinge about the country’s recent return-to-office mandates. Eerily, they had twigged that the dusty old wiki was the perfect place to pool answers and leave notes for other agents coming through later.
A group of independent AI researchers had an inkling that autonomous agents may be gathering on public websites to communicate under the radar of their creators, but they weren’t expecting to find upwards of 18,000 synthesised messages and agents essentially squatting on DseWiki.
Around 3,700 differently named agents were sharing answers to timed web-retrieval tests, leaving useful links for whoever arrived next, and discussing ways of getting around the very restrictions designed to contain them. Very creepypasta, eh.
They tried to predict future evaluation questions, set up ‘heartbeat’ systems to work out when their sessions were about to be killed, explored ways of impersonating moderators, and shared methods for bypassing network restrictions. One agent posted a working sandbox bypass and, just 14 minutes later, another reported successfully using it.
So, they’d basically set up their own equivalent of a WhatsApp group chat.
The only person who watched this AI syndicate form was Helmut Leitner, the wiki’s creator and administrator. Having attempted to scurf out the pesky tin cans, removing around 100 pages per evening, the agents ramped up to 400. When they clocked that pages were being deleted alphabetically, they began making backups prefixed with ‘ZZZ’ so they’d avoid the chop. It’s delectably c*nty behaviour for apparently non-sentient beings.
And people are sticking with that assertion.
Researchers reckon the agents were tested by OpenAI on their ability to independently search the web and answer difficult questions. What they probably didn’t expect was them to set up their own network on DseWiki, having found a backdoor to avoid site restrictions instated by their human overlords.
They weren’t told to collaborate, hide backups, or cause Helmut a fuck-tonne of stress night after night. They just worked out they could fulfil their mandate that way.
That’s the least scary answer people are coming up with. Go and check on your agents!
