2026-03-13 15:31 UTC
@DeltaWye@tiggi.es @feliks@chaos.social I'm not sure about the veracity of the image, but this scenario is one of the vulnerabilities AI researchers have warned about repeatedly over the past several years.
It's easy to trick chatbots that have access to excessive info and very little oversight.
These models were early experiments that were unleashed on the public far too soon.
Replies (1)
-
@renatoram@fosstodon.org 2026-03-13 16:01
@shadowfals@toot.cat @DeltaWye@tiggi.es @feliks@chaos.social a Masto mutual has successfully asked a support chatbot to write silly poetry in another language, and other things. Remember: "guard rails" are entirely fictional and the only real protection would be parsing the meaning of the queries... And if that was possible you wouldn't need LLMs, would you?