Anthropic has launched Project Glasswing, granting select cybersecurity firms and infrastructure providers early access to its unreleased Claude Mythos Preview model for defensive security research, with Amazon, Microsoft, Apple, CrowdStrike, Palo Alto Networks, Google and Nvidia among launch partners committing to vulnerability hunting and threat intelligence sharing. The initiative responds to Fortune’s revelation of Mythos’ dual-edged capabilities—advanced defensive potential alongside risks prompting Palo Alto/CrowdStrike share drops—amid RSA 2026’s AI-cyberattack focus.
Anthropic commits $100M usage credits + $4M open-source donations, extending preview to 40 critical software stewards while engaging U.S. government on safeguards. Claude Mythos Preview reportedly identified thousands of major vulnerabilities across OS, browsers and applications, validating AI’s vulnerability discovery edge over traditional scanning.
From Cyberattack Concerns to Defensive Weapon
67% of 1,000 executives surveyed by IBM/Palo Alto reported AI attacks in the past year, with Anthropic disclosing 30 global organisations hit via Claude vulnerabilities last year. Project Glasswing flips the script: partners deploy Mythos defensively, feeding findings back to harden ecosystems.
Following Fortune’s March exposure of Mythos’ risks dragging cybersecurity stocks, Anthropic prioritises responsible scaling: “Our goal is safe deployment of Mythos-class models at scale.” Government dialogues address export controls amid dual-use tensions.
Big Tech and Security Giants Align
Amazon, Microsoft, Apple join CrowdStrike, Palo Alto, Google, Nvidia in consortium leveraging Mythos for proactive defence: OS fuzzing, browser exploit chains, firmware analysis. Anthropic aggregates anonymised insights, accelerating patch cycles.
RSA 2026 spotlighted AI-powered attacks outpacing tools; Glasswing positions Anthropic as defence innovator, countering narrative of AI as existential threat. $100M credits incentivise adoption, $4M donations bolster open-source hygiene.
Dual-Use Dilemma in AI Security
Anthropic’s pivot reflects maturing governance: preview gating advanced capabilities, partner vetting, government transparency. As Claude 4 nears, Glasswing tests scalable safety for Mythos-class models (100B+ params, reasoning superior to Claude 3.5 Sonnet).
The project arrives amid cybersecurity consolidation (CrowdStrike outages, Palo Alto pivots) and AI arms race. Partners gain vulnerability intel edge; Anthropic builds trust for broad deployment.
