Post

Eh, I tend to agree with last year's report by IBM Red that, once the LLM OPS attack surfaces become profitable, there will be a big problem there. Carlini claims "hundreds" of vulns in the Linux kernel he hasn't reported yet. That's coherent long-term project with clear reporting system. +

If there were a way to find and then chain together vulns on MCP + a couple other things, it might get bad. I think the biggest potential effect Mythos-level LLMs might have is that attack surfaces become profitable sooner.

I work front lines. There's a lot of concern about the accumulated risk from back level systems and potential in change of pace of exploitation. So far, the change has been in self-injury from agentic experiments. This will be a step change like fuzzing was on early introduction.

Concern is different to use 😅