Post

Mark Histed
@markhisted.org

Yes, they figured out how to use cache timing to communicate across VMs—but humans figured this out a long time ago. The Spectre/Meltdown side channel attacks in 2018 were cache timing leaks. And they were widely written up. So the LLMs had a record of cache attacks in training data. 1/3

Brendan Nyhan@brendannyhan.bsky.social

holy moly openai.com/index/securi...

Mark Histed
@markhisted.org

(In a previous life I was a contractor for the MIT computer security team. But it’s been a long time. Others working in the field at present may chime in.)

Mark Histed
@markhisted.org

Cc @prasad.bsky.social and @joshtpm.bsky.social who I saw amplify this. It’s not that it’s not impressive, it is. But these kind of timing (and maybe even temperature?! I vaguely remember this) sidechanneling effects across processes have been widely discussed and worked on by humans.

0 reposts7 likes

Prasad Jallepalli, MD, PhD
@prasad.bsky.social

as I alluded to, it’s the same principle as American POWs in Hanoi tapping code to each other across a concrete wall regardless of how it was implemented in this specific case, the general type of solution is something we know about from human history and experience

Mark Histed
@markhisted.org

Yep, it’s a weird kind of side channel but one that’s not totally unfamiliar. I went off in reaction to people who seemed to be saying this was revolutionary for AI (not you) but maybe I misread them