Mythos Didn't Change the Game. It Put It on a Schedule.

While I was off the grid last week, the vulnerability landscape shifted under all of us.

I was away with my family, with limited internet access (and it was wonderful). I came back to a pile of unanswered pings asking for my thoughts on Mythos, Glasswing, and whether the sky is falling. Short answer: the sky isn’t falling. It’s been descending slowly since 2024, and some of us have been pointing it out and getting politely ignored.

Two years ago, foundational models could take a CVE write-up and generate exploit code. We knew model-generated code was riddled with security issues from day one. The trajectory was always headed here; Mythos didn’t invent the problem, it put it on a schedule, and it is not the finish line. I fully expect other foundational models, if they’re not already there, to match or exceed these capabilities. This is an industry trajectory, and the capability floor for vulnerability discovery and exploitation is dropping across the board.

Sixty-plus contributors wrote a rapid-response “Mythos-Ready” briefing via CSA/SANS. It’s outstanding: a real risk register, built over a weekend because CISOs needed it Monday morning, not when it was perfect. I don’t challenge a word in it.

But here’s my less comfortable addition: we should have been ready for this already. I spent years at FS-ISAC making this exact case to the financial sector and co-authored what was arguably the first comprehensive industry paper on cryptographic agility in financial services. Patching discipline, crypto-agility, and designing for rapid-response updates: none of these are new ideas. They’re just ideas that kept losing budget fights to shinier things. Every year we deferred that work, we were writing IOUs, and Mythos is the collections department calling. With USG officials reportedly holding closed-door briefings with bank CEOs about AI-driven cyber risk, this is no longer a conversation that lives in the SOC. It’s a systemic stability issue.

Now zoom out. Earlier this year, I wrote that agentic coding tools like Claude Code are fundamentally changing the rate of software production. That prediction is confirmed. GitHub’s COO recently shared that the platform went from 1 billion commits in all of 2025 to 275 million per week, on pace for 14 billion this year. That’s a 14x increase in output volume. AI agent-initiated pull requests went from 4 million last September to over 17 million in March. Connect the dots: we’re generating code at 14x the volume, and now we have models that can find and exploit vulnerabilities in that code faster than any human team. The math is not subtle, and we need to fundamentally rethink how we deliver and support software. The old cadence of applying patches and annual pen tests was already outdated; now it’s dangerous.

One genuinely encouraging note: in Anthropic’s own testing, Linux kernel mitigations held. Mythos found the bugs but couldn’t exploit them remotely. Defense in depth, the boring, unsexy, conference-talk-nobody-attends kind of work, is what held the line.

This is also an inflection point in how we engage early-career professionals, and I think the connection matters. For years, we’ve onboarded junior talent by teaching them the patterns of the past and having them earn their stripes through bug fixes and maintenance. That’s still important; you need to understand why those patterns exist. With agentic tools democratizing code production and AI-powered vulnerability discovery compressing attack timelines, we also need fresh thinkers who can leapfrog the old playbook. Early-career professionals who grew up with these tools may be the ones who figure out how to build systems that are secure by default at this new speed, not because they learned our habits, but because they never inherited our blind spots. We should be pulling them into the conversation now, not asking them to wait their turn.

The takeaway isn’t panic; it’s urgency. Stop treating patching and hardening like the vegetables on your plate and start treating them like the oxygen in your lungs. The window between disclosure and weaponization has just compressed from weeks to hours. Your response cadence needs to match.