Anthropic opens Project Glasswing to 150 critical-infrastructure operators in 15 countries

The Mythos Preview expansion lands a day after Anthropic confidentially filed for an IPO, and pulls power, water, hospitals and telecoms into the cohort for the first time.

Anthropic on Tuesday extended Project Glasswing to roughly 150 new organizations across more than 15 countries, pulling power utilities, water systems, hospitals, and telecoms into a cohort that until April had been confined to 50 hyperscaler-and-bank partners running the Claude Mythos Preview model. The expansion arrived exactly one day after the company confidentially filed an S-1 with the SEC, a sequencing decision that’s doing real work.

The original cohort, per Anthropic’s blog post, surfaced more than 10,000 high- or critical-severity vulnerabilities. CyberScoop reports that several partners told Anthropic their bug-discovery rates rose more than tenfold after deploying the model. Those numbers are the throughline of the IPO narrative: this isn’t a research preview, it’s a deployed capability with measurable industrial output across AWS, Cisco, CrowdStrike, Google, JPMorgan Chase, the Linux Foundation, Microsoft, Nvidia, Palo Alto Networks, Apple, and the U.S. government.

Rubrik confirmed its inclusion in the new wave. Cybersecurity Dive reports ENISA, the EU’s cybersecurity agency, has also been admitted. Anthropic withheld the rest.

Use cases now extend past vulnerability discovery into patch authoring, pre-release checks, penetration testing, and translating legacy code into memory-safe languages. Anthropic frames the present bottleneck not as finding flaws but as “verifying, disclosing, and patching” them, a position echoed by a joint Cloud Security Alliance, SANS Institute, and OWASP report cited by CyberScoop warning that defenders are likely to be outpaced near-term by AI-equipped attackers.

The financial frame is hard to ignore. TechCrunch pegs Anthropic’s most recent round at $65 billion on roughly a $1 trillion valuation; CNBC framed the S-1 as beating OpenAI, which has rolled out GPT-5.5-Cyber to a large testing group, to public markets. Anthropic says it won’t ship a general-release Mythos-class model until safeguards exist, pointing retail customers to Claude Security on Opus 4.8 instead, and estimates rivals are 6 to 12 months from parity.

A capability demonstration, a customer list, and an S-1, in that order. The TARP-era playbook of fact-on-the-ground before paperwork, ported to frontier AI.