Anthropic Expands Claude Mythos Access in 2026

TL;DR

Anthropic is expanding Project Glasswing from 50 to approximately 200 partners, granting access to Claude Mythos Preview—a model more powerful than the public Opus family—specifically for defensive cybersecurity work
Partners have already identified over 10,000 high- or critical-severity vulnerabilities across major operating systems, browsers, and critical infrastructure codebases
The expansion includes underrepresented sectors like power, water, healthcare, and communications across 15+ countries, targeting infrastructure that affects hundreds of millions of people
Industry critics argue Anthropic’s validation process lacks transparency, with concerns that contractor-based review doesn’t meet open peer-review standards the security community expects

What Happened

Anthropic announced Tuesday it’s adding roughly 150 organizations to Project Glasswing, the initiative that provides controlled access to Claude Mythos Preview for vulnerability detection. The original cohort of 50 partners—including AWS, Apple, Google, Microsoft, and the Linux Foundation—launched in early April 2026.

The new partners span industries previously underrepresented in the initial group: power grids, water utilities, healthcare systems, telecommunications, and hardware manufacturers. Many are vendors whose codebases serve as dependencies for thousands of other organizations globally, including government agencies.

Anthropic’s warning was stark: “A successful attack on their codebase could be catastrophic. For most partners, we estimate that a major attack could affect more than 100 million people, with important ramifications for both global and national security.”

The company justified the expansion by pointing to AI’s current capabilities: “AI models have reached a level of coding capability where they can surpass all but the most skilled humans at finding and exploiting software vulnerabilities.” Anthropic predicts that within 6-12 months, multiple AI companies will have Mythos-class models—potentially without safeguards.

Why It Matters

This is a pre-emptive defense against a capability race that’s already underway. OpenAI released GPT-5.5-Cyber through its Trusted Access for Cyber program in May 2026, with scaled deployment following in mid-April. The offensive potential of these models creates asymmetric risk: defenders need months to patch vulnerabilities that offensive models could exploit in hours.

The scale of discovered vulnerabilities validates Anthropic’s concern. Over 10,000 high- or critical-severity flaws found across partner codebases in just weeks suggests the attack surface is far larger than traditional security audits have revealed. When a single vulnerability in a widely-used library can cascade across millions of systems, the multiplier effect of AI-accelerated exploitation becomes existential.

For developers and security teams, this marks a fundamental shift in the vulnerability disclosure lifecycle. The bottleneck is no longer discovery—it’s verification, prioritization, patching, and deployment at a scale that human processes weren’t designed to handle.

Key Details

Claude Mythos Preview capabilities:

More powerful than the publicly available Opus family of models
Already identified vulnerabilities in “every major operating system and web browser”
Can be used for penetration testing, automated threat detection, and legacy code migration to memory-safe languages

Project Glasswing structure:

Original 50 partners (April 2026)
Approximately 150 new partners added (May 2026)
Partners span 15+ countries
Each organization must meet security requirements before access
Access includes proprietary tools developed for vulnerability detection

Supporting infrastructure:

Claude Security (released February 2026): Public service using Claude Opus 4.8 to scan codebases and suggest patches
Partners share best practices and triage findings with third parties
Anthropic provides vulnerability disclosure frameworks for open-source maintainers

Implications

Anthropic is betting that controlled proliferation of defensive AI beats the alternative: offensive capabilities reaching threat actors first. This approach assumes responsible disclosure and coordinated patching can outpace weaponization.

The geographic and sectoral expansion reveals where Anthropic sees the highest risk: critical infrastructure. Power, water, healthcare, and communications systems weren’t built with modern cybersecurity assumptions. Many run legacy code where a single compromise could have cascading physical-world consequences.

The 6-12 month timeline matters. Once Mythos-class capabilities become commoditized, the window for securing vulnerable systems closes. The expansion signals Anthropic believes that window is narrowing faster than initially projected.

Our Take

Anthropic is making the right strategic move, but the execution has a legitimacy problem.

Justin Beals of Strike Graph nailed the core issue: “That’s not third-party validation, that’s editing.” When Anthropic selects which vulnerabilities get independently reviewed and hires the reviewers, it’s self-grading its homework. For a model this consequential, the security community deserves full-corpus, independent evaluation—not curated highlights.

The peer-reviewed standards Beals references exist for a reason. Software trust is built on verifiable claims, not corporate assurances. As Mythos-class models become infrastructure dependencies, opaque validation processes create technical debt that compounds with every integration.

That said, doing nothing isn’t an option. Guy Currier of Futurum Group is correct that “something broad-based has to be done, and the sooner the better.” The criticism should push Anthropic toward transparency, not paralysis.

Watch for three things:

Whether competing labs follow OpenAI’s “lockstep” approach of scaling defensive capabilities with model releases, or fragment into incompatible security frameworks
How fast the 10,000+ vulnerabilities get patched versus how quickly offensive capabilities proliferate beyond trusted access programs
Whether Anthropic opens its validation process to independent security researchers, or whether external pressure from partners forces the issue

The cybersecurity equilibrium just shifted. The question is whether defenders can patch faster than attackers can adapt.