Anthropic Project Glasswing: 150 Partners and 10,000 Vulnerabilities Later

Ten thousand serious software vulnerabilities. That’s what Anthropic’s roughly 50 existing Project Glasswing partners have surfaced to date — and the program hasn’t reached its intended scale yet.

What Happened

Anthropic is expanding Project Glasswing to approximately 150 new partner organizations, bringing the total to around 200 across more than 15 countries. The program uses Claude Mythos Preview — Anthropic’s specialized security-focused model — to scan critical software for vulnerabilities. New partners span energy, water, healthcare, and communications infrastructure: sectors where a successful attack could affect more than 100 million people.

In parallel, Anthropic launched Claude Security, a commercial product built on Claude Opus 4.8 that performs codebase scanning and patch recommendations for enterprise customers. The EU’s cybersecurity agency ENISA is reportedly negotiating access terms for the program.

Why It Matters

The 10,000 vulnerability figure matters because it’s not a marketing claim — it’s an inventory of unpatched attack surface that existed in critical infrastructure before AI-assisted scanning surfaced it. Traditional security audits are expensive, slow, and constrained by the supply of qualified researchers. AI scanning doesn’t replace that expertise, but it dramatically expands the coverage area per unit of time.

What Glasswing represents is a shift in who owns the security research bottleneck. If AI models can autonomously triage codebases for vulnerability classes — buffer overflows, injection vectors, authentication bypasses — the constraint shifts from “finding researchers” to “deploying and interpreting agents at scale.” That’s a different problem, and one where Anthropic has an obvious infrastructure interest.

Anthropic’s note that rival companies “could ship comparable models within 6 to 12 months, potentially without safeguards” reads as both a competitive warning and a justification for Glasswing’s intentionally gated rollout. Claude Mythos is not generally available — a deliberate choice to prevent dual-use at scale before protective measures are established. The EU’s ENISA negotiating access reflects the geopolitical dimension: AI-powered vulnerability scanning is inherently dual-use infrastructure.

What This Means for Builders

Security teams in critical infrastructure should engage with the Glasswing program now, before it closes to new partners. The combination of Claude Mythos for research-grade scanning and Claude Security (Opus 4.8) for commercial codebase review suggests Anthropic is building a two-tier model: purpose-built access for infrastructure defenders, commercial access for enterprise engineering teams.

For developers building on the Claude API, the emergence of a security-specialized model variant is a directional signal. Expect purpose-built model families — security, legal, medical — to become the default architecture for regulated-industry deployments. General-purpose models will handle general tasks; specialized variants will handle liability-sensitive ones.

Source: The Decoder

Content created with AI assistance and reviewed for accuracy.