Anthropic Previews Claude Mythos and Launches Project Glasswing — a $100M Controlled Security Rollout
Anthropic has previewed Claude Mythos, a tier above Opus, and simultaneously launched Project Glasswing — a controlled early-access program giving 11 major tech partners and 40+ organizations priority access specifically to use Mythos for finding and fixing security vulnerabilities. Anthropic is backing the rollout with $100M in model credits and $4M in open-source security donations.
Original sourceAnthropic today unveiled two interconnected announcements: a preview of Claude Mythos, its next frontier model, and the launch of Project Glasswing, an unprecedented controlled rollout designed to harness Mythos's cybersecurity capabilities before general availability.
**What is Claude Mythos?**
Mythos is positioned as a step above Claude Opus — Anthropic's existing top-tier model — with claimed improvements across coding, reasoning, and in particular cybersecurity research. The system card notes that Mythos "surpasses all but the most skilled human experts at finding exploits in software," making it one of the most capable offensive security tools ever created. This capability is precisely why Anthropic chose not to launch it broadly.
**Project Glasswing: Defense First**
Rather than releasing Mythos to all users simultaneously, Anthropic structured a controlled early-access program: 11 major technology partners (including AWS, Apple, Google, Microsoft, and NVIDIA) plus 40+ vetted organizations get priority access — but primarily for defensive security use. The framing is "use this powerful tool to find and fix vulnerabilities before bad actors find and exploit them."
Anthropic is backing the program with $100 million in model credits for participants and $4 million in direct donations to open-source security projects. The structure is an implicit acknowledgment that deploying a superhuman vulnerability-finder requires a community commitment to responsible use.
**Industry Reaction**
The announcement dominated Hacker News with three separate threads accumulating over 650 upvotes combined. The most common sentiment: this is the correct sequencing — capability + controlled deployment + defensive focus — but many observers noted the program creates a two-tier access model where well-resourced companies get Mythos first, potentially widening the security gap between large and small organizations. The $4M open-source donation was seen as meaningful but insufficient given the scale of the capability.
Panel Takes
The Builder
Developer Perspective
“The controlled rollout model makes sense for a model this capable — but the 11 major partners getting first access means startups and independent researchers are again at a disadvantage. Anthropic should have included a 'small team' tier in Project Glasswing from day one.”
The Skeptic
Reality Check
“Project Glasswing is smart PR, but 'controlled access' programs consistently expand faster than intended and the controlled period rarely lasts long. The real question is what Anthropic's liability looks like when Mythos's offensive capabilities are eventually used for harm — and they will be.”
The Futurist
Big Picture
“Anthropic has just established a new template for deploying dual-use AI: capability preview → curated defensive rollout → community investment → general release. Whether this template holds as AI capabilities scale further is the defining governance question of the next five years.”