@flyingpenguin@infosec.exchange
Post #4364150
2026-08-03 21:07 UTC
@Viss@mastodon.social @lorenzofb@infosec.exchange exactly. even worse is we have the precedent in actual engineering history for this exact problem and the prevention. it's so simple.
here's my deck, which I taught many, many CS graduate students and even corporations with. my experience with the OpenAI and Anthropic crew is they think their wheel isn't like any other wheel because they believe in hallucinations more than history.
https://www.flyingpenguin.com/2015-fachhochschule-st-polten-ethical-foundations-for-information-security/
Replies (1)
-
@flyingpenguin@infosec.exchange 2026-08-03 21:11
@Viss@mastodon.social @lorenzofb@infosec.exchange and for what it's worth I just saw someone with 60K followers post their "agentic sandbox" spec and it has an obvious egress flaw. like, AI (computer science really) has no actual integrity breach detection. you can call a four sided box sealed and there's no agency to declare it false.