“This supports a bit of a theory that I've had, which is that bugs that are discovered with LLMs are essentially public, because it's almost guaranteed that there's going to be parallel discovery.”
Risky Business · Software Engineering & Infra · November 2026
Context
A researcher had just said Microsoft patched a bug he found with an OpenAI model before he could report it, because others had found the same flaw. Patrick uses this to state his theory that LLM-found vulnerabilities are effectively public.
Transcript
Now, when you were dealing with, was it you dealing with Microsoft directly on this?
…manual research on and off. And then around the second week of September, I found it. And then I applied the second patch that came out on September third week, I think. And it was patched. So we did not even get a chance to report the vulnerability to Microsoft. Yeah, right, right, right. Because
they found it as well, right? When we're doing the same thing. This is interesting though, right? Because this has come along, like this, this supports like a bit of a theory that I've had, which is that bugs that are discovered with LLMs are essentially public because there is almost, it's almost guaranteed that there's going to be parallel discovery. I mean, would you take that as a lesson from this?
Yep, that was a big lesson from this. So when SMB was patched, Elad Shamir, the VP of research at SpectroOps, challenged me to go and look after the LDAP as well because the same bug might exist on LLAP side too, and we can bypass the LLIP signing. So that's when I pivoted from SMB to LDAP and I…
Yep. And they patched this one as well.
So we reported this one to Microsoft in the first week of October. And a patch came out in the second week of October. And when we published a blog post and a tweet on X, right, for this research, another research reached out to me and said they reported this to Microsoft two months ago. And their…
More from Risky Business
