Anthropic opens Project Glasswing to ~200 organizations with Claude Mythos Preview

• AnthropicAI's expansion post says Project Glasswing is growing by about 150 organizations, taking Claude Mythos Preview from roughly 50 initial partners to about 200 vetted organizations across more than 15 countries.
• According to Anthropic's announcement, the new cohort leans toward critical infrastructure and upstream software vendors, while cedric_chee's note highlights that Anthropic is prioritizing maintainers and nonprofits whose code reaches far beyond a single enterprise.
• Anthropic's initial update says the first 50 partners found more than 10,000 high- or critical-severity flaws, and Anthropic's disclosure dashboard shows why the hard part has shifted from finding bugs to validating, disclosing, and patching them.
• Anthropic's Mythos technical assessment says the model can identify and exploit zero-days, chain vulnerabilities, and help non-experts build working exploits, which is why rohanpaul_ai's summary thread describes Mythos as closer to a cyber weapon detector than a normal coding assistant.
• The expansion still stops well short of general release: Anthropic's announcement says Mythos-level access needs stronger misuse safeguards first, while TheRealAdamG's BBC link post points to OpenAI offering a rival cyber model to UK banks that Anthropic has not admitted to Mythos previews.

You can read the expansion post, browse the Glasswing program page, and dig through Anthropic's public disclosure dashboard to see how much of the bottleneck has moved from discovery to remediation. The weirdest detail lives in Anthropic's Mythos capability writeup, which describes a model that can turn subtle bugs into working exploits across major operating systems and browsers. Meanwhile the BBC report linked by TheRealAdamG shows the rollout is already geopolitical, with OpenAI pitching GPT-5.5 Cyber to UK banks that still do not have Mythos access.

Anthropic's June 2 update is simple on paper: about 150 new organizations are being added on top of the roughly 50 partners announced in April, with security checks before each organization gets access.

The official post says the new group spans more than 15 countries and fills gaps in sectors that were underrepresented in the first cohort, including power, water, healthcare, communications, and hardware. Anthropic also says most of these organizations provide critical infrastructure to many more countries, and that for most partners a successful attack could affect more than 100 million people.

The original Project Glasswing page had a much more tech-industry-heavy launch roster, including AWS, Apple, Google, Microsoft, Cloudflare, CrowdStrike, the Linux Foundation, and JPMorganChase. This expansion is the first real signal that Glasswing is moving from big-name launch partners toward broader infrastructure coverage.

Anthropic's most interesting filter is not company size. It is software leverage.

The announcement says many of the new partners are vendors, meaning companies or nonprofits that maintain codebases relied on by other organizations, including governments. That matches rohanpaul_ai's summary thread, which describes the new list as a defensive priority set that includes software maintainers and nonprofits alongside operators in power, healthcare, water, communications, and government.

That matters because Glasswing is not being framed as a premium scanner for whoever can pay. Anthropic is picking codebases where one patch can travel downstream to huge numbers of systems, which is a very different access policy from ordinary enterprise model rollouts.

Anthropic's own reporting now treats bug discovery as the solved part of the loop.

In Anthropic's initial update, the company says the first month of Glasswing produced more than 10,000 high- or critical-severity findings, with several partners reporting bug-finding rates more than 10 times higher than before. The public coordinated vulnerability disclosure dashboard makes the next bottleneck visible: as of May 22, Anthropic had disclosed 1,596 vulnerabilities across 281 open-source projects, but only 97 were patched.

The June 2 post pushes the same idea harder. Anthropic says partners are already using Mythos to write patches, run pre-release checks, support penetration testing, automate threat detection and response, and even help rebuild legacy systems in memory-safe languages. The pitch has shifted from finding more bugs to increasing the throughput of verification, disclosure, and repair.

Mythos stays gated because Anthropic is describing something much stronger than a code reviewer.

In Anthropic's technical assessment of Mythos Preview, the company says the model can identify and exploit zero-day vulnerabilities in every major operating system and web browser it tested. The writeup claims Mythos chained four browser bugs into a sandbox escape, found a now-patched 27-year-old OpenBSD bug, and helped Anthropic engineers without formal security training wake up to complete working exploits.

That same report says over 99% of the vulnerabilities Anthropic found were still unpatched at publication time, which is why the company is keeping Mythos in a tightly controlled preview instead of opening normal API access. Anthropic's June 2 post goes further and says it expects other labs to have Mythos-class cyber models within 6 to 12 months, potentially without the same safeguards.

The last new detail in the expansion post is that Anthropic is quietly building access layers below full Mythos release.

Anthropic says it recently launched Claude Security, which uses public models such as Opus 4.8 for defensive scanning and patch suggestions, and that it is also releasing some of its Glasswing-developed tools on request to trusted security teams. The same post says Anthropic plans to scale up a Cyber Verification Program that would grant Mythos-class capabilities for specific cyberdefense tasks before general access is safe.

That staged rollout helps explain the competitive split now forming around these systems. As the BBC report linked by TheRealAdamG notes, OpenAI has already offered GPT-5.5 Cyber to several major UK banks that Anthropic has kept out of Mythos previews, while Anthropic is still expanding through vetted cohorts and task-specific verification rather than open availability.