Anthropic Just Turned AI Into a Security Team
Anthropic announced Project Glasswing this week, a collaborative security initiative with 11 major technology partners: AWS, Apple, Broadcom, Cisco, CrowdStrike, Google, JPMorganChase, Linux Foundation, Microsoft, NVIDIA, and Palo Alto Networks. The goal is to find and fix critical software vulnerabilities before malicious actors do.
The model powering it, Claude Mythos Preview, found a 27-year-old vulnerability in OpenBSD, a 16-year-old flaw in FFmpeg that survived five million automated fuzzing runs, and multiple chained Linux kernel vulnerabilities enabling privilege escalation. These weren't new codebases with obvious mistakes. They were mature, battle-tested systems reviewed by thousands of engineers over decades.
The standout number isn't the age of the bugs. It's the five million automated tests on the FFmpeg flaw. Every existing security tool missed it. The AI didn't.
What the model is actually doing
Claude Mythos Preview doesn't just scan for known patterns. It develops exploits autonomously to verify whether a flaw is exploitable. It scores 83.1% on CyberGym, compared to 66.6% for the previous best. That's a jump from "useful assistant" to "autonomous security researcher."
To find a 27-year-old vulnerability in OpenBSD, the model had to hold a mental model of the entire system, trace data flows through code it didn't write, and construct a chain of conditions to prove exploitability. That's reasoning previously exclusive to senior security researchers.
The same capability transfers to programming broadly. A model that can reason about exploit chains can reason about architectural flaws, performance bottlenecks, and API design. Security just produces the most legible proof.
The coalition is the story
Eleven companies, direct competitors in many cases, agreed to share access to a pre-release AI model, pool resources, and commit $100M in credits toward finding bugs in shared infrastructure. Apple and Google at the same table. Cisco and CrowdStrike, both in.
They're willing because open-source infrastructure is shared infrastructure. OpenBSD, FFmpeg, the Linux kernel run inside every company's stack. A vulnerability there is a vulnerability everywhere. And the maintainers of that critical software often work voluntarily, with no budget for security audits. Glasswing is trying to fix that gap at scale.
What it signals
The window between a vulnerability being discovered and exploited "has collapsed," as the CrowdStrike CTO put it. Glasswing is a bet that AI can close that gap on the defensive side. Finding vulnerabilities used to require expensive, scarce specialists. An AI system that does this across millions of lines of code simultaneously changes the economics.
Claude Mythos Preview will cost $25 per million input tokens and $125 per million output tokens. Expensive, but cheaper than a three-month security audit.
The pressure will move through the industry in layers. Security review cadences that felt adequate in 2024 will look negligent in 2026. If AI can audit for security at this depth, it can audit for correctness and performance too. "Did you do a security audit?" will become "did you run AI-assisted vulnerability analysis?" Teams that can answer confidently will have an edge.
What this means if you're building software
The immediate impact is indirect. Glasswing's patches will flow through your normal update cycles. The longer-term question is harder to ignore: if AI can find 27-year-old vulnerabilities in code that millions of expert eyes reviewed, what's sitting in code that hasn't been reviewed nearly as carefully?
Most production codebases have never had a serious security audit. The right response is treating security as continuous, not something scheduled for a future sprint. The same AI tools that help you write code faster can find what you missed.
The gap between teams using AI for security analysis and teams not using it will widen fast. Same pattern as automated testing and CI/CD. Early adopters look paranoid, then there's an incident, then everyone adopts. The teams that wait pay more.
Glasswing is a proof of concept for something bigger. The real question isn't what Anthropic found in OpenBSD. It's what the same capability looks like pointed at a typical startup's production codebase.
A project named after transparency
The glasswing butterfly (*Greta oto*) has transparent wings. The symbolism: transparency that makes hidden things visible. Vulnerabilities hiding in plain sight in code everyone depends on.
The 27-year-old OpenBSD vulnerability didn't appear this week. It's been there since 1999. The question was always whether anyone would find it before the wrong person did.