„The paper cites a race condition that an agent attempted to fix by inserting a delay in the code with a Thread.Sleep“
Meaning, it was in its training data.
Meaning, for better agents you need b...
@nprofile1q... This, exactly, why I don't see how current LLMs (statically trained) can ever produce secure code - since there's so little of it in the training data and you can't really tag it as such either since secure code snippet + secure code snippet well can become unsecure code snippet.
We would need a new paradigm where LLMs can retrain themselves continously as humans do and over time learn how to write better code.
I'm sure it'll happen, but we're not there.
We would need a new paradigm where LLMs can retrain themselves continously as humans do and over time learn how to write better code.
I'm sure it'll happen, but we're not there.