5
The Triage Pipeline and the "Flood" Problem 13:47 Nia: I’m thinking back to what we said about the "flood" of vulnerabilities. If Mythos finds ten thousand bugs, how do you actually disclose them without just breaking the internet? I mean, if every major browser gets a thousand bug reports on the same Tuesday, the developers are just going to quit.
14:03 Eli: That is a very real fear in the community. If you automate discovery but don't automate triage and disclosure, you just create a massive bottleneck. Anthropic’s Newton Cheng, who leads their frontier red team, addressed this directly. He said they built a specific "triage pipeline" to manage this. They don't just dump the raw output on a maintainer. They have a system where a "secondary" AI agent confirms the significance of the bug first.
14:25 Nia: Oh, so the AI checks the AI's work?
1:53 Eli: Exactly. And then, for high-severity stuff, they actually have professional human triagers they’ve contracted to manually validate the bug before it gets sent out. They are trying to ensure that every report a maintainer gets is high-quality and actionable. They even said they won't submit large volumes to a single project without talking to the maintainers first to agree on a pace they can handle.
14:48 Nia: That sounds responsible, but it also sounds slow. If the "window" between discovery and exploitation is collapsing to minutes, can we really afford to wait for a human triager to double-check everything?
15:01 Eli: That’s the million-dollar question. Anthropic is trying to follow a "coordinated vulnerability disclosure" framework—basically, they give the developer 90 days to fix it before they go public. But as CrowdStrike’s CTO pointed out, the adversaries aren't going to follow that 90-day rule. If an attacker has their own "Mythos-class" model, they’re going from discovery to exploit in minutes.
15:22 Nia: So the "good guys" are following the old rules of etiquette while the "bad guys" are playing at warp speed. That’s a terrifying asymmetry.
2:37 Eli: It really is. And that’s why Anthropic is experimenting with "candidate patches." When they report a bug, they try to include a patch written by the AI. They label it as "model-written" so the maintainer knows to scrutinize it, but it gives them a head start. It’s not just—"Hey, you have a problem"—it’s—"Hey, you have a problem, and here’s a possible fix."
15:49 Nia: That’s a huge value-add. But I remember seeing something in the sources about "patch quality." You can't just blind-trust an AI to write a kernel patch, right?
15:58 Eli: Definitely not. Newton Cheng was very clear about that—autonomously written patches need the same level of testing and scrutiny as human ones. But even if the patch is only 90 percent of the way there, it’s a massive time-saver for a stressed-out developer.
16:12 Nia: I also noticed they’re using cryptographic hashes for the bugs that haven't been patched yet. What’s the point of that?
16:17 Eli: That’s a way of "proving" they found it without "leaking" the details. They publish a hash today, and then 90 days later, when they reveal the full bug, that hash proves they were the first to find it. It builds trust in the research process and prevents "claim jumping" in the security world. It’s about being a "responsible citizen" in the vulnerability research space.
16:35 Nia: It’s like they’re trying to build a new set of "norms" for the AI era. But as you said, norms only work if everyone agrees to them. If a state-sponsored group from, say, North Korea or Russia gets a similar model, they aren't going to be publishing hashes and waiting 90 days.
1:53 Eli: Exactly. Anthropic actually noted that a Chinese state-sponsored group was already using their older models for "autonomous tactical execution" back in 2025. They were hitting 30 targets, including tech firms and government agencies. And that was with the *old* stuff. So the sense of urgency in Project Glasswing isn't theoretical. It’s based on the fact that the "offensive" use is already happening.
17:14 Nia: That really puts the "Glasswing" name in a different light. It’s not just about camouflaging vulnerabilities—it’s about the fact that we are all living in a glass house now. If the AI can see through all the layers of code, there’s nowhere left to hide.
17:28 Eli: That’s a chilling thought. But it’s also why the collaboration between rivals like Google and Microsoft is so significant. They realize they’re in the same glass house. If the underlying infrastructure of the internet—the Linux kernel, the web browsers, the cloud protocols—is compromised, it doesn't matter who has the better "features." The whole business model of the digital economy depends on trust, and that trust is what’s at stake here.