Anthropic Expands Project Glasswing to 150 New Organizations and Launches Claude Security: What It Means for Your API Routing Policy

Anthropic's Project Glasswing—their collaboration with critical-infrastructure organizations to use Claude for vulnerability scanning—has entered a new phase. On June 2, Anthropic announced that the program is expanding from its initial cohort of approximately 50 partners to roughly 150 new organizations, alongside the public launch of Claude Security, a new product built on Claude Opus 4.8.

What happened

The initial Glasswing cohort, operational since early April, has already uncovered more than 10,000 high- or critical-severity security flaws across partner codebases using Claude Mythos Preview. The expansion adds partners in 15+ countries, targeting industries not well represented in the initial round—power, water, healthcare, communications, and hardware—and many of the new partners are vendors whose codebases are depended upon by governments and other large organizations worldwide.

Anthropic estimates that a successful attack on most new partners' codebases could affect more than 100 million people.

Alongside the expansion, Anthropic launched Claude Security, a product that uses Opus 4.8 to scan codebases and suggest patches. This is the first time Anthropic has packaged a frontier model specifically for security use as a standalone product.

Why it matters for routing teams

1. Claude Security is now a real product tier

Claude Security is not an internal tool or research preview—it's a released product that uses Opus 4.8 as its backbone. For teams routing API traffic to Anthropic, this means:

• Opus 4.8 is the security-recommended model. Anthropic's own security product uses it, not Sonnet or Haiku. If your organization has compliance or audit requirements that touch code scanning, routing those workloads through Opus 4.8 is now the Anthropic-blessed path.
• New billing surface. Claude Security pricing isn't yet disclosed on the API changelog, but operators should expect it to carry Opus-class pricing or higher for codebase scans. If your routing policy uses tiered models based on task sensitivity, security scanning now has an explicit product recommendation.

2. Mythos Preview capabilities set the 6–12 month baseline

Anthropic explicitly stated in the announcement:

"Within 6 to 12 months, we expect that many other AI companies will have Mythos-class models, and they could release them without safeguards that prevent misuse."

For routing teams, this is a concrete timeline signal. If your policy currently treats frontier reasoning models as optional for security workloads, you now have a vendor-stated window for when comparable capabilities will be widely available—likely across multiple providers.

What to do now: Audit whether your routing policy has a security-scanning tier. If it doesn't, Anthropic's own model recommendation (Opus 4.8) and the 6–12 month proliferation timeline give you a defensible starting point.

3. The vulnerability-finding → patching pipeline is bottlenecked by human review

Anthropic's own announcement identifies the bottleneck: Mythos-class models can surface vulnerabilities faster than organizations can verify, disclose, and patch them. Claude Security is designed to help with this—it writes patches, not just reports.

For API operators, this means the value of routing security workloads to Claude is not just detection—it's end-to-end remediation. If your pipeline currently uses one model for detection and another for patch generation, the Glasswing results suggest a single-model pipeline (Opus 4.8) is Anthropic's recommended architecture.

What didn't change

• No new model announced. Mythos Preview remains the same model. Claude Opus 4.8 remains the generally available backbone.
• No API pricing changes tied to this announcement. The expansion is about product reach, not API surface.
• No new routing primitives. Claude Security is a product layer, not a new API mode. Standard Messages API routing applies.

What to watch next

• Claude Security pricing and API surface. If Anthropic exposes Claude Security as an API-modeled product (rather than a dashboard-only tool), routing teams will need to add it to their cost models.
• Mythos Preview GA timeline. If Mythos moves from preview to general availability, it will become a new model tier in routing policies—and potentially a new cost ceiling.
• Competitor security models. Anthropic's "6–12 months" estimate implies OpenAI, Google, and others will ship comparable capabilities. Multi-provider security routing will become a real architectural decision.

Bottom line

Project Glasswing's expansion from 50 to 150 organizations—combined with the launch of Claude Security as a product on Opus 4.8—cements Anthropic's position as the first frontier model provider to ship a dedicated security-scanning product. For routing teams, the signal is clear: security workloads now have an explicit model recommendation, and the 6–12 month proliferation window for Mythos-class capabilities means security-aware routing policies are no longer optional—they're on a deadline.