Jon Green

speaker
145 appearances 1 recordings 1 series first heard Aug 2026 last heard 6 Aug

Jon Green’s voice in public audio — every appearance, attributed to the second.

Trend

recordings per month · last 12 months
1 · Aug OctJan 26AprJulnow

Recordings per month over the last 12 months — 1 in all, peaking in Aug 2026 with 1.

Appearances

newest first · ▶ plays the moment
Project Glasswing is an initiative started by Anthropic, obviously a frontier AI model lab.
What they discovered was in roughly April of 2026, their latest AI model, which they called Mythos, had been trained really well on software engineering.
And as a side effect of being good at software engineering, it actually was also very good at security vulnerability analysis when you give it source code.
And they were a little scared by how good it was that a large language model was finding and chaining together security vulnerabilities from products that have been out for, in some cases, 30 years or more.
And so what they said is, and some of this is marketing, of course, but they said this model is too powerful to release to the public right now.
However, it has such potential for making the cybersecurity world better.
We want to give early access to critical infrastructure providers and people who make critical infrastructure to try to get ahead of their own vulnerabilities in their products.
And so they were kind enough to invite us into that project.
And the primary thesis behind that was around Juniper, the fact that Juniper powers a good portion of the public internet.
And that's where we got started with that.
So this has been about six weeks old at this point.
And so it's sort of obvious, but why is that so important?
I've become convinced that software engineers probably cannot write secure code in the face of multi-threaded applications with low down to the hardware level memory management.
It's not because they're not good engineers.
It's because the problem space is extremely difficult.
We've known this for a long time.
So just having these things out there and knowing that AI models are good at finding them
If the wrong people get a hold of those models and analyze those products, it's very bad for all of us.
So this is basically using an AI model to sort of pen test software?
It really is.
Showing 1–20 of 145 · page 1 of 8 Next →