Ariel Herbert-Voss

We found an 0day in CPython that turns your LLM's chat template into a code execution primitive. A backdoored model on Hugging Face can own your inference server 🧵

“Patch everything” is easy advice to give when you’re not the one keeping everything running. If you’re the one responding to everything happening this week, here’s how I’d put agents to work 🧵:

jeez I go heads down for a few hours and the whole world explodes

What today's stories make even more clear is how much of our security depends on attacking being expensive. AI reduces the effort needed to turn a difficult bug into an exploit, and it helps a hacker pursue more targets at scale.

I'm still optimistic about defense, but that optimism has a major caveat: we have to make protection scale. Yes, finding the issue first is valuable. But then we need to identify who and what is exposed, deliver a fix or containment measure, and verify it worked. That's the only way defenders should measure progress: how quickly we close real opportunities for attackers

I spoke to @josephmenn this week about AI worms and wanted to add more color to my thoughts. What makes an AI worm different is that a model is capable of learning on the fly. This type of threat isn't here yet, but this is an example of an increasing shift that defenders are feeling.

We keep hearing the same theme from @runsybil customers: finding vulns is only the start. Teams need to patch and understand their attack surface in real time. Our job is to find and fix what's exploitable faster than attackers. Onward 🚀

Full article here: theguardian.com/technology/202…

it’s an insane time to be building in security

we closed out Q3 at @runsybil last week as our strongest quarter yet. 10x revenue growth over Q2

I’m really proud of the team we’ve built to meet this moment