Mastodon Feed
cwebber@social.coop ("Christine Lemmer-Webber") wrote:
If OpenAI got the info on techniques for mathematical advancements from private interactions being trained into the model, then I wonder: could the Hugging Face and some other recent attacks also have *partly* (not entirely) used private information?
Specifically, OpenAI said that the models got information on API keys "from a pasteboard" somewhere, which is possible but something felt off about that.
Now I realize what it is: if someone was using an agent running *on their computer* which *had access to their secrets*, that information could have made it into the training set! Which means, the agents could have "known" about certain useful private API keys.
And in fact, this may be happening all over the place with coding agent use...