Matthew Green
@matthew_d_green
I teach cryptography at Johns Hopkins. Mostly on BlueSky these days at
1.1K Following 158.7K Followers
So one thing I’ve heard from multiple researchers in my field is that models seem to get *much* better at solving their specific problems over short periods, even asked in new contexts. Many of them have mentioned that they wonder if they’re training on their inputs.
Show more
More TEE attacks, this time on Signal’s contact discovery enclaves.
Oh boy, we had better get our security act together. And fast.
Something we need to start talking about urgently: huge numbers of old software bugs are being found and closed right now. Intelligence and law enforcement agencies depend on these bugs for access. What are they going to do when this access is gone?
Show more
Hey, these folks did it! Pulling encrypted reasoning out of frontier models using cross-model replays. I wonder if Anthropic and OpenAI care now?
It’s very hard to distinguish AI security incidents from AI marketing, and that’s actually a big problem going forward.
There’s been some reporting that Meta contributed an unfathomable sum to promote age verification laws globally. This is broadly true, but actual situation is a bit more complex. Figured it was worth an update.
Show more
