Anthropic expands Claude Mythos cyber model to ~150 more organizations under Project Glasswing

Anthropic on June 2, 2026 said it would extend access to its powerful, not-publicly-released Claude Mythos model to roughly 150 additional organizations across more than 15 countries under Project Glasswing, its initiative to harden critical software and infrastructure. The expansion brings the total to about 200 participants and broadens coverage into sectors underrepresented in the initial cohort—power, water, communications, healthcare and hardware. Anthropic said the roughly 50 initial partners, who began deploying Mythos Preview in early April to scan their codebases, have surfaced more than 10,000 high- or critical-severity security vulnerabilities.

The company framed the program around defensive asymmetry: many new partners are vendors whose code underpins widely used systems, and Anthropic estimated that for most partners a major attack could affect more than 100 million people, with national- and global-security implications. The move is significant because Mythos remains withheld from general commercial release on safety grounds—Anthropic's April 244-page system card described it as simultaneously its best-aligned model and the one posing the greatest alignment-related risk, and independent evaluators (UK AISI) found Mythos-class capability among the first to complete an end-to-end simulated corporate network intrusion. Distributing such a capability to a controlled set of defenders, rather than the public, mirrors the logic of the Trump administration's June 2 executive order favoring defender-first access to frontier cyber capabilities (filed same day).

Anthropic positioned the expansion as preventing catastrophic compromise of essential software while keeping the underlying model gated.