Claude Mythos Goes Global: Inside Anthropic’s Project Glasswing Expansion

Anthropic's Claude Mythos Preview is not just another AI model; it's a controlled cybersecurity weapon that the company deemed too dangerous to release publicly. Instead, Anthropic created Project Glasswing, a tightly restricted initiative that partners with select organizations to deploy the model defensively—finding and fixing critical software vulnerabilities at machine speed. After a successful initial phase, the project is now undergoing its largest expansion, moving from U.S. tech giants into the core infrastructure of the global economy.

From 50 to ~200: The Biggest Project Glasswing Expansion Yet

In early June 2026, Anthropic announced a massive scale-up of Project Glasswing, extending access to its Claude Mythos Preview model from around 50 initial partners to approximately 150 new organizations . This brings the total participant count to around 200 organizations spanning more than 15 countries .

The newly added cohort marks a strategic pivot from the initial group of major tech firms and cybersecurity companies—which included Amazon Web Services, Apple, Google, Microsoft, and CrowdStrike —to a much broader set of operators running critical infrastructure. The expansion now reaches deep into sectors like power, water, healthcare, telecommunications, and hardware manufacturing .

The geographic footprint has also widened dramatically. While the initial phase was largely limited to the U.S. and UK, the new partner list spans the Five Eyes intelligence alliance (Canada, Australia, New Zealand) alongside major economies including France, Germany, Japan, South Korea, Spain, and India . To support this global push, Anthropic also opened a new office in Milan, Italy, its sixth European location .

Organizations do not receive automatic access. Each new partner must pass specific security checks and requirements defined by the program before gaining entry .

The NYSE Joins the Hunt: Securing Global Financial Plumbing

Perhaps the most significant new partner is Intercontinental Exchange (ICE), the parent company of the New York Stock Exchange. On June 3, 2026, ICE announced it had joined Project Glasswing and is deploying Claude Mythos Preview across its extensive global infrastructure .

This deployment covers not only the iconic stock exchange but also ICE's clearinghouses, data services, and mortgage technology platforms . According to ICE President Ben Jackson, the goal is to identify and remediate security vulnerabilities before they can be exploited by emerging AI-powered threats, thereby strengthening the "security and resilience of critical market infrastructure" .

The Scorecard: 10,000+ Zero-Days and a 73% Expert-Level Success Rate

The expansion is backed by staggering performance numbers that illustrate why organizations are clamoring for access. In just its first month of operation through late May 2026, Claude Mythos Preview discovered more than 10,000 high- or critical-severity zero-day vulnerabilities across widely deployed software .

The model's offensive capabilities were rigorously tested by the UK AI Security Institute (AISI). In its evaluation, AISI reported that Mythos Preview achieved a 73% success rate on expert-level capture-the-flag (CTF) cybersecurity tasks—challenges that no prior AI model could complete before April 2025 . AISI also observed that when given network access, Mythos Preview could autonomously discover and exploit vulnerabilities, executing multi-stage attacks that would take human professionals days of work .

In a practical demonstration of its power, cybersecurity firm Palo Alto Networks used Mythos to identify 75 valid vulnerabilities. This aligns with broader validation data showing that of the high- or critical-severity issues flagged by Mythos and assessed by independent firms, 90.6% were confirmed as true positives .

The Open-Source Vulnerability Firehose and a $100 Million Pledge

Anthropic didn't just scan partner codebases; it also pointed Mythos at open-source software, unleashing a torrent of responsible disclosures. As of May 22, 2026, Anthropic had formally disclosed 1,596 vulnerabilities across 281 open-source projects .

Of those disclosures, 97 had been patched, and 88 were assigned formal CVE records or GitHub Security Advisories (GHSA) . In many other cases, project maintainers shipped fixes without issuing a public advisory. To support maintainers facing this sudden workload, Anthropic pledged up to $100 million in cloud credits to help them fix the uncovered flaws .

The raw numbers were even larger. Mythos Preview scanned over 1,000 open-source projects and estimated a total of 23,019 vulnerabilities, including 6,202 deemed high or critical severity. Of the high/critical issues that underwent human triage, 90.6% were confirmed valid .

A Model Too Risky to Release Publicly—For Now

Despite the rapid expansion of Project Glasswing, Anthropic’s stance on public release remains firm: Claude Mythos Preview is not and will not be released to the general public, citing fundamental safety and misuse concerns .

The model can autonomously identify and exploit zero-day vulnerabilities in every major operating system and web browser . It generated 181 working browser exploits where a previous model managed only two—a 90x increase—and successfully turned Firefox vulnerabilities into working exploits 72.4% of the time .

Anthropic has indicated a path toward eventually releasing future "Mythos-class" models with added safeguards, but no public release date has been set. For the moment, the model remains accessible exclusively through Project Glasswing, available via the Claude API, Amazon Bedrock, Google Cloud's Vertex AI, and Microsoft Foundry for authorized users .

White House Pushback and Resolution

The road to this massive expansion wasn't entirely smooth. In late April 2026, the White House opposed an earlier, smaller Mythos expansion plan over national security concerns . The administration and Anthropic engaged in several weeks of close collaboration that also included the security industry, open-source maintainers, and the U.S. government. The differences were ultimately resolved ahead of the larger rollout announced on June 2–3, allowing it to proceed .

Defending at Machine Speed: Cisco’s 1.8 Billion Lines of Code

Within Project Glasswing, Cisco demonstrated the raw speed of AI-powered defense. The company used the model to scan 1.8 billion lines of code in just 8 weeks, uncovering vulnerabilities at a pace no human team could match. As a direct result of this accelerated detection capability, Cisco has moved to a twice-monthly vulnerability disclosure cadence .

Europe’s Response: BNP Paribas Doubles Down on Mistral AI

The staggering power of models like Mythos is also reshaping the global competitive landscape. In direct response to the cybersecurity threat posed by such frontier AI, BNP Paribas has expanded its partnership with Mistral AI. The French bank is increasing its investment and deployment of Mistral's models to build sovereign European AI-driven defenses, explicitly framing it as a critical infrastructure protection strategy against models developed by U.S. firms like Anthropic .