What Just Happened
Anthropic built their most powerful AI model ever then refused to release it.
They called it Claude Mythos Preview.
Instead of launching publicly, they created Project Glasswing a closed group of 40+ organizations given controlled access for one purpose only: defense.
Why This Model Is Different
This is not a smarter chatbot. It operates at a different level entirely:
Finds zero day vulnerabilities in every major OS and browser
Chains exploits autonomously, with minimal human guidance
Discovered bugs hiding for 27 years that every prior audit missed
Scores 93.9% on software engineering benchmarks vs 80.8% for the previous best
It performs like an elite security researcher. At machine speed. Around the clock.
Why Anthropic Refused to Release It
Their position is direct:
Offensive cyber capability at this level scales instantly
Attackers move faster than defenders can patch
Critical infrastructure hospitals, grids, banks could be exposed before anyone responds
AI has now surpassed most humans at hacking, making broad release too risky.
Anthropic CEO Dario Amodei
Their word for it: a cybersecurity reckoning.
Project Glasswing (The Defensive Move)
Rather than release Mythos publicly, Anthropic handed limited access to a trusted coalition:
Who's in:
Apple, Google, Microsoft, AWS
CrowdStrike, Palo Alto Networks, Cisco
Linux Foundation + 35 other infrastructure groups
What they're getting:
Up to $100M in API credits
Direct funding for open source security patches
Early access to scan and fix vulnerabilities before bad actors find them
The goal: patch the world's critical software before comparable AI capabilities spread to the wrong hands.
What's Already Happening
This isn't theoretical. It's live:
Security teams are scanning large legacy systems right now
Financial institutions have started accelerated patch cycles
Open source maintainers are running high priority audits
The clock is already ticking.
A Voice From the Front Line
"Claude Mythos demonstrates what is now possible for defenders and adversaries will inevitably look to exploit the same capabilities." Elia Zaitsev, CTO, CrowdStrike
The tension is real: the same tool that defends can attack.
What Changes From Here
Three things are now becoming clear:
AI is an active cyber weapon : not theoretical, not future, operational today
Selective release becomes normal : some models may never be made public
The arms race is live : comparable systems are likely already in development elsewhere
Bottom Line
It's proof that AI capabilities have crossed a threshold where who has access matters as much as what the model can do.
The bug hid for 27 years. An AI found it in minutes.
The question now is who builds the next one, and what do they do with it?
Related Articles and Sources
Official Anthropic: Project Glasswing announcement and technical preview details.
System Card: Detailed 244-page report on capabilities and safety (available via Anthropic links).
CBS News: Coverage on vulnerabilities and risks.
Fortune (leak context): Early details from March 2026.
TechCrunch: Preview and initiative summary.
Found this useful? Forward it to someone in tech. They need to know.