Anthropic Brings Its Most Powerful Mythos AI Model to India; Expands Project Glasswing To 14 Other Countries

Anthropic has announced the largest expansion yet of Project Glasswing, the company's restricted cybersecurity initiative built around its most advanced, and publicly unavailable, AI model, Claude Mythos Preview. The San Francisco-based AI company said it is bringing approximately 150 new organisations on board, more than tripling the size of the programme, which initially launched in early April with roughly 50 partners.

India is among the new countries receiving access, alongside Canada, Australia, France, Germany, Japan, and South Korea, according to an Anthropic blog. The company said organisations in the new cohort span more than 15 nations, with most maintaining software and infrastructure that serves users and governments far beyond their home countries.

The Financial Times reported that the expanded cohort includes technology companies Samsung, SK Telecom, and SK Hynix; cybersecurity firm Okta; financial messaging network Swift; settlement platform Euroclear; and the Intercontinental Exchange, which operates the New York Stock Exchange. Defence and government bodies NATO and the European Union Agency for Cybersecurity (ENISA) are also said to be among the new partners.

Anthropic said the new group covers industries that were underrepresented in its original cohort, including power, water, healthcare, communications, and hardware, sectors that sit at the heart of national critical infrastructure. Many of the new entrants are vendors, organisations whose codebases are relied upon by other companies, governments, and millions of end users downstream.

Anthropic's latest cybersecurity push

Anthropic did not mince words about why these organisations were selected. The company said that for most of the new partners, a successful cyberattack on their systems could affect more than 100 million people, with significant consequences for both national and global security. That framing underscores the rationale for keeping Mythos Preview restricted, and why the company believes giving these organisations early access to the model's capabilities is a defensive imperative.

What does Claude Mythos Preview do, and why it remains locked away?

Claude Mythos Preview is Anthropic's most capable model, built around advanced coding and reasoning abilities that allow it to analyse large codebases and surface security vulnerabilities that have evaded detection, sometimes for decades. The company has chosen not to release it publicly, citing the model's ability to not just identify but also construct exploits for the flaws it finds.

In an earlier update on Project Glasswing's progress published in May, Anthropic said the roughly 50 initial partners had collectively found more than 10,000 high or critical-severity vulnerabilities in the weeks since the programme launched. Cloudflare alone, one of the original partners, found 2,000 bugs across its critical-path systems, 400 of which were rated high- or critical-severity. The company said Mythos Preview was the first AI model to solve both of the UK's AI Security Institute's cyber range simulations end to end.

Anthropic says that the rollout is well thought out

Anthropic acknowledged in its announcement that the expansion is partly driven by urgency. The company said it expects that within six to twelve months, other AI companies will develop models with comparable cybersecurity capabilities to Mythos Preview, and that those companies may release them without equivalent safeguards. The risk, Anthropic argues, is that the same power being used today to find vulnerabilities could soon be cheaply deployed by attackers to exploit them.

The company said it sees its role as twofold: helping the software industry adapt by safely widening access to its most capable models, and shifting the programme's focus over time from finding flaws to actively supporting their disclosure, patching, and deployment. Alongside the Glasswing expansion, Anthropic said it has released Claude Security, a product built on its publicly available models, to Enterprise customers, and is making internal scanning tools available to qualifying security teams on request.