// news · alignment · agents2026-05-27source: anthropic / alignment.anthropic.com / dark reading

Project Glasswing uses Claude Mythos preview to help secure world's most critical software — Anthropic frames defensive-cyber as alignment priority

Project Glasswing's expanded scope this cycle frames defensive cyber explicitly as an alignment priority — using the Claude Mythos preview to help secure the world's most critical software and prepare industry for keeping ahead of cyber attackers. The framing matters because it positions cyber capability as a safety-relevant axis the lab will invest in proactively, rather than as a capability tier the lab merely restricts.

The framing shift is the strategic substance. Through 2024-2025 the dominant lab posture toward cyber capability was restriction — refuse-list training, deployment-policy controls, exception processes for legitimate security research access. The Glasswing framing inverts that posture: cyber capability is recognized as one of the safety-relevant axes where defender capability needs to lead attacker capability, and the lab commits to investing in defender tooling proactively. The Mythos preview is the most-capable Claude model approved for cyber work, and Glasswing is the deployment surface that makes it available to vetted defenders.

The practical scope covers the critical-software defense problem. Open-source software supply chains, infrastructure-as-code repositories, security-critical libraries and frameworks, software-bill-of-materials attestation — these are the defensive surfaces the program targets. Combined with the Claude Security public beta's SOC-tooling reach and the cyber-verification tooling for eligible security teams, Glasswing now spans the prevention, detection, and response surfaces simultaneously. The alignment-relevant claim is that defender lead-time is itself a safety property — and one the lab's capability advantage uniquely lets it improve.

See our analysis →

Anthropic — Project Glasswing scope and Mythos preview deployment → · Anthropic Alignment — Mythos preview defensive-cyber research notes → · Dark Reading — AI defender lead-time 2026 commentary →