Why we're watching this: Anthropic is moving Claude Mythos out of a controlled pilot and into live critical infrastructure across 15 countries, with named partners including NATO, Samsung, and the EU's cyber agency. How this scales over the next six to twelve months will set the baseline for how AI gets used in national security infrastructure.
Key Takeaways
- 150 new organizations across more than 15 countries now have access to Claude Mythos Preview through Project Glasswing
- Partners include NATO, Samsung, SK Hynix, SK Telecom, Okta, and ENISA, covering power, water, healthcare, communications, and hardware sectors
- Anthropic estimates a successful attack on most partners’ codebases could affect more than 100 million people
- Initial 50 partners already found more than 10,000 high- or critical-severity vulnerabilities since April
- Rival OpenAI has launched GPT-5.5-Cyber, now in partner testing, signaling direct competition in AI-powered cybersecurity
What Happened
Anthropic is expanding Project Glasswing, its AI-powered vulnerability detection program, to roughly 150 new organizations across more than 15 countries, the company announced Tuesday, one day after filing confidentially for an IPO.
The new cohort covers sectors not well-represented in Anthropic’s initial April group, including power, water, healthcare, communications, and hardware. According to a person familiar with the matter cited by the Financial Times, the expanded group includes organizations in Australia, Canada, France, Germany, Italy, Switzerland, the Netherlands, Spain, Belgium, Sweden, India, Japan, New Zealand, and South Korea.
Partners named by the FT include U.S. identity management firm Okta, South Korean companies Samsung, SK Hynix, and SK Telecom, NATO, and the EU’s cybersecurity agency ENISA. The common thread across all new partners, Anthropic said, is exposure to catastrophic risk. For most partners, the company estimates a major attack could affect more than 100 million people.
Claude Mythos Preview sits at the center of Project Glasswing. In the roughly six weeks since the initial 50 partners gained access, those organizations collectively found more than 10,000 high- or critical-severity vulnerabilities. Cloudflare found 2,000 bugs across its critical-path systems, with 400 rated high or critical severity. Mozilla found and fixed 271 vulnerabilities in Firefox 150, more than ten times the number found in Firefox 148 using Claude Opus 4.6.
Why It Matters
The expansion lands alongside direct competitive pressure. OpenAI has released GPT-5.5-Cyber, its own cybersecurity-focused model, currently in partner testing. The parallel rollouts signal that AI-powered vulnerability detection is fast becoming a standard enterprise security offering rather than a niche government program. Anthropic said it plans to continue expanding Project Glasswing, with future cohorts targeting additional essential infrastructure providers and maintainers of critical open-source software.
Critics and independent analysts have pointed out that a program this concentrated carries its own risks. Anthropic itself acknowledged that no AI company, including Anthropic, has yet developed safeguards strong enough to release Mythos-class capabilities to the general public without meaningful risk of misuse. The company says it expects Mythos-class models to be widely available from other AI providers within 6 to 12 months, potentially without equivalent safeguards in place.
“What each partner has in common is that a successful attack on their codebase could be catastrophic.” — Anthropic, Project Glasswing expansion announcement
Anthropic has not yet developed safeguards it considers strong enough to prevent Mythos from being misused at general release scale. The company says it expects Mythos-class models to be widely available from other AI providers within 6 to 12 months, potentially without equivalent safeguards.
Mythos Preview and similar models are now surfacing vulnerabilities far faster than security teams can triage, patch, and deploy fixes. Anthropic’s own data shows a steep drop-off at each phase of its disclosure pipeline. Teams adopting AI vulnerability scanning should have a clear patching workflow before they start scanning, not after.
Bottom Line
Watch whether OpenAI’s GPT-5.5-Cyber moves from partner testing to broader availability before Anthropic scales its Cyber Verification Program. Whichever lab establishes the dominant enterprise security partnership network first will be significantly harder to displace, given the depth of access these programs require into production codebases.
For security and ops teams, the Glasswing results set a new baseline expectation for AI-assisted vulnerability scanning: thousands of findings in weeks, not months. Teams that have not yet built a patching and triage workflow capable of absorbing that volume should treat that gap as the priority before evaluating any AI scanning tool. Relve is an AI trends intelligence platform covering how deployments like Project Glasswing are reshaping enterprise security posture and the competitive dynamics between frontier AI labs.
