Anthropic Project Glasswing Expands to 150+ Organizations in 15+ Countries Including NATO and ENISA

Project Glasswing expands from ~50 to 200+ partner organizations across 15+ countries

The second wave adds roughly 150 organizations. Named participants include Okta, Samsung, SK Hynix, SK Telecom, NATO, and the EU cybersecurity agency ENISA. Countries covered include Australia, Canada, France, Germany, Italy, Switzerland, Netherlands, Spain, Belgium, Sweden, India, Japan, New Zealand, and South Korea — all U.S.-aligned nations.

Expansion deliberately targets critical infrastructure sectors absent from the initial cohort

Power, water, healthcare, communications, and hardware were underrepresented in the first 50 partners. The new cohort focuses on organizations whose codebases underpin systems that other companies and governments depend on — maintainers of shared infrastructure rather than individual enterprises.

For most partners, Anthropic estimates a major attack on their codebase could affect more than 100 million people

This is Anthropic's stated criterion for inclusion in the expanded cohort. The company explicitly frames the stakes in terms of both national security and global security implications, not purely commercial risk.

Initial 50-partner cohort found 10,000+ critical vulnerabilities within the program's first month

The scale of discovery in a single month underscores the step-change in throughput that AI-assisted vulnerability research represents. The bottleneck has since shifted from finding flaws to verifying, disclosing, and patching them at comparable speed.

Cloudflare found 2,000 bugs — 400 high or critical — with false positive rates better than human testers

Low false positive rates are operationally significant: high false positive rates cause security teams to deprioritize AI-generated findings. Cloudflare's results suggest Mythos Preview is producing actionable output, not noise.

Mozilla found 271 vulnerabilities in Firefox 150, more than 10x the count achieved with Claude Opus 4.6

The Opus 4.6 comparison provides a direct generational benchmark. A 10x uplift in a single model generation on the same codebase signals nonlinear capability growth in this domain.

UK AI Security Institute confirmed Mythos Preview is the first model to complete both cyber range exercises end-to-end

The cyber ranges test a model's ability to execute multistep cyberattacks autonomously. Government validation that a commercial model can complete these exercises marks a threshold: frontier AI offensive capability is now operationally real.

OpenAI released GPT-5.5-Cyber, a competing cybersecurity model now in large-scale partner testing

Anthropic had publicly anticipated that rival models of comparable capability would arrive quickly. GPT-5.5-Cyber's release validates that prediction and adds urgency to Glasswing's mission of establishing defensive norms before such tools proliferate broadly.

Anthropic filed confidentially for an IPO following a $65B funding round at a nearly $1 trillion valuation

The funding round and IPO filing were disclosed the day before the Glasswing expansion announcement. At a ~$1 trillion valuation, Anthropic would rank among the most valuable technology companies at the time of any public offering.

Claude Mythos Preview is Anthropic's most powerful model, purpose-evaluated for offensive and defensive cybersecurity

Released in early April 2026 specifically in the context of Project Glasswing. Its ability to identify thousands of zero-day vulnerabilities over several weeks, combined with UK AISI end-to-end cyber range results, distinguishes it from prior general-purpose models in this domain.

Anthropic is racing to build disclosure and patching infrastructure before Mythos-level capability becomes widely available

The company has framed Glasswing's urgency explicitly around proliferation risk: once comparable models are in broad circulation, the defensive advantage of controlled access disappears. The expansion to allied-nation critical infrastructure is partly an attempt to institutionalize secure disclosure norms while a window remains.

The structural bottleneck in software security has shifted from vulnerability discovery to verification, disclosure, and patching

This framing emerged from Glasswing's first-month results and is now being tested at scale across international critical infrastructure. Traditional security team workflows built around human-paced discovery are no longer the rate-limiting constraint.