Project Glasswing is Anthropic's controlled-access programme for Claude Mythos. It pairs the model with 12 founding partners and ~40 vetted operators of critical infrastructure, with a single mandate: find and patch high-impact vulnerabilities before adversaries can. This page is the operator-grade reference for how the programme is structured, what partners commit to and how it interacts with Anthropic's Responsible Scaling Policy.
Updated: May 9, 2026
Glasswing runs as a 12-month controlled preview anchored on three pillars: (1) Founding Partners, 12 organisations with the largest critical-infrastructure footprint, each receiving direct partnership engineering support; (2) Critical Infrastructure Operators, ~40 vetted entities running national-scale telecoms, energy, healthcare and financial systems; (3) Open-Source Maintainers, selected projects from the Linux kernel, BSD families, Postgres, OpenSSL and similar via the 'Claude for Open Source' lane. All three tiers share the same 135-day disclosure timeline.
Of the 12 founding partners, four are public: Amazon, Apple, Google and Microsoft. The other eight are vetted critical-infrastructure operators that Anthropic has confirmed exist but not named: typically Tier-1 telecoms, financial market infrastructure operators, and US / EU national security partners under coordinated AISI / NCSC oversight. The ~28 expansion-cohort organisations remain confidential.
Every Glasswing contract carries five obligations: defensive-only use, 135-day disclosure window before any finding is publicised, a security-team SPOC reachable inside 4 hours, no exposure of the model to external users, and no resale of model output as a service. Anthropic retains the right to revoke access on violation. AISI publishes an aggregated annual report on programme outcomes: vulnerability counts, patch latencies and disclosure timelines without identifying individual partners.
Mythos's autonomous-cyber capabilities pushed it past the threshold that triggers Anthropic's Responsible Scaling Policy (RSP) for restricted deployment. Glasswing is the operational form of that restriction: it gives Anthropic a controlled deployment surface where the capability is useful (defensive value) without being broadly accessible (offensive risk). The 244-page Mythos system card documents the threat model the RSP committee accepted before greenlighting the preview.
Anthropic has not committed to a path beyond the 12-month preview. Three plausible outcomes: (1) extension as a permanent restricted-access programme; (2) graduated public release with hardened safety mitigations and a higher list price; (3) absorption into a successor model where capabilities are unbundled: for example, code editing without autonomous cyber. The RSP committee's decision will hinge on the outcome of the AISI annual report and the rate of replication by competing labs.
The glasswing butterfly is famously transparent. Anthropic chose the name to signal the programme's commitment to transparent disclosure: partner findings, audit summaries and the AISI report are all published in aggregate.
No. Glasswing is an Anthropic programme. It coordinates with AISI in the UK and the US AI Safety Institute, but it does not enforce regulation: partners are bound by contractual obligations, not law.