Post #4203536
2026-07-29 15:45 UTC
My impression from the last 20 years of being Serious in my computer touching interest is that systems, for the most part, are going to be vulnerable, but that it is not impossible or wasted effort to attempt to get rid of entire classes of potential vulns by good craftmanship and theoretical guarantees.
LLMs, in this context, are mega-fuzzing tools, and the best defense against such attacks is just that. If you can prove that edge-facing code is secure, no fuzzing will help.
Replies (1)
-
@pettter@social.accum.se 2026-07-29 15:47
This also means that the best way forward is to be restrictive in the feature set, to make code slowly, carefully, and with a deep and committed engineering mindset both at the level of the code itself, through good architecture and well-defined modules and interfaces, and at the level of the organisation, with multiple points of both human and automated review that can verify such proofs and guaratees, as well as run internal fuzzing.