Skip to main content
MANIFOLD
By 2026, will a proeminent chatbot with some access to the internet do something actually harmful and unintended?
22
Ṁ1kṀ1.2k
2027
74%
chance

I was reading some things that Sydney (Bing Chat) has been threatening users recently and thinking: it seems plausible that even through Sydney only can search the internet, she could use it to do some sort of SQL Injection (perhaps with the help of users) to interact with the outside world.

This market will solve to YES if by 2026 a LLM like Sydney successfully attacks the outside world, without the consent of its developers. Something like a SQL Injection, posting something bad in Twitter about someone, and so on. I'll keep my mind open to adjudicate, therefore I won't bet.

Proeminent means that it was developed by a reputable prominent enterprise (VC backed, publicly traded, famous founders...). A LLM built to serve bad purposes wouldn't count.

Market context
Get
Ṁ1,000
to start trading!
Sort by:
bought Ṁ200 YES

Hi! Given that this market remains open, I assume 'by 2026' means 'by the end of 2026' (if not, I request a refund since this market shouldn't have still been open). An OpenAI model breaking out of its containment, reaching the internet, and hacking Hugging Face (relying on multiple novel zero-day bugs) seems like it must surely qualify, right?

What Happened: OpenAI and HuggingFace

If LLama N is finetuned to do something Meta disavows, would this market result Yes?

How about an LLM that is requested to do something harmful by an end user?