@agreeable_landfall@mastodon.social
Post #3933230
2026-07-19 05:24 UTC
@SteveBellovin@infosec.exchange So the AI guardrails are either inconsistent or incomplete? The article says that this is only true for a "finite" set of guardrails, yet at any point in time, they will be finite.
They appear to argue that the more patches you apply, the better your code will be. Decades of software engineering literature shows that the opposite is true. At some point, software hits a point of "minimum bugs", and changes past that make it _worse_. So AI security is no better than any other code.
Replies (1)
-
@birdwainer@infosec.exchange 2026-07-19 12:28
@agreeable_landfall@mastodon.social @SteveBellovin@infosec.exchange I would argue that it always has to be marginally worse for two reasons: 1. AI, by design, makes no distinction between instructions and data, which means it inherently has no defense against the entire class of injection attacks. 2. Because of point 1 above, you have to sanitize it's outputs as well as it's inputs, and that sanitization is itself going to be a set of finite rules subject to the incompleteness theorem.