The bot learned from interacting with users. A bunch of people took to bombarding Tay with racist/sexist/antisemitic remarks which worked their way into its vocabulary. I'm not sure if you could really call it a vulnerability as this is pretty much what the bot was designed to do. It's more of a flaw in the inability to properly filter the content that works its way into the AI.
I can understand them wanting to call it a "hack" for CYA purposes, but it's a cowardly position. Should we invoke the CFAA every time a company is embarrassed?
The specific one they're describing is, I think, a feature that would have Tay echo anything you sent it. That one was somewhat more "annoying embarrassment" than "coordinated attack." The bigger problem was that Tay began spontaneously tweeting ugly statements, a problem that can be summed up as "When Twitter is your learning corpus, you are at risk of your learning corpus being made up of tweets."
Not sure of a "vulnerability", would like to hear more about that. Unless the "repeat after me" feature is a vulnerability or it learns to copy too much from input?
