In what may be one of the most significant cybersecurity milestones of the decade, Anthropic has revealed the initial results of Project Glasswing — a collaborative AI-driven vulnerability research initiative that used its unreleased Claude Mythos Preview model to autonomously discover more than 10,000 high- and critical-severity zero-day vulnerabilities across some of the world’s most widely deployed software systems.
What Is Project Glasswing?
Project Glasswing is a defensive cybersecurity consortium involving Anthropic and over 50 major technology organizations, including Microsoft, Apple, Google, and Cloudflare. The initiative was designed with a single mission: deploy advanced AI to find and remediate critical vulnerabilities in critical infrastructure before malicious actors can exploit them.
The project leveraged Claude Mythos Preview — an AI model withheld from public release due to severe dual-use risks — to autonomously scan highly targeted codebases. The model demonstrated an unprecedented ability not only to identify complex security flaws, but also to construct functional exploit code entirely on its own.
Staggering Scale of Discovery
The numbers from the first month of operation are extraordinary. Over 10,000 high- and critical-severity zero-day vulnerabilities were identified across partner codebases. Cloudflare alone reported 2,000 bugs found, including 400 rated high or critical severity, with the model’s false-positive rate outperforming human security testers. Mozilla used Claude Mythos Preview to uncover and patch 271 vulnerabilities in Firefox 150 — ten times more findings than previous testing with Claude Opus 4.6.
The UK’s AI Security Institute confirmed that Mythos Preview is the first model to fully solve its multistep cyberattack simulations. In open-source scanning across more than 1,000 widely used projects, a notable discovery was CVE-2026-5194, a critical flaw in the wolfSSL cryptography library that allowed for the forgery of security certificates — enabling attackers to spoof banking or email domains invisibly.
The Vulnerability Management Crisis
The Glasswing results have exposed a critical structural weakness in the global software ecosystem: human capacity to triage, report, and patch vulnerabilities cannot keep pace with AI-driven discovery. The initial scanning phase produced 23,019 candidate findings. When 1,900 were reviewed by external security firms, 1,726 (90.8%) were confirmed as valid true positives. Despite Anthropic reporting 1,596 vetted findings directly to maintainers, only 97 vulnerabilities have been patched upstream — a massive remediation gap leaving millions of users exposed.
This bottleneck highlights the severe capacity constraints faced by volunteer open-source maintainers now overwhelmed by high-quality AI vulnerability disclosures.
A New Era of Cyber Risk
The industry is entering a transitional phase that fundamentally changes the threat landscape. Because Mythos-class models reduce the cost and time of zero-day discovery to nearly zero, the traditional 90-day coordinated vulnerability disclosure window now poses new dangers. The lag between AI discovery and widespread patch deployment creates a highly exploitable window for threat actors who may develop or acquire similar capabilities.
Due to these dual-use risks, Anthropic has withheld Mythos from public release entirely, restricting its use exclusively to defensive consortium members.
Supporting the Broader Ecosystem
To help the wider security community while Mythos remains restricted, Anthropic has launched Claude Security in public beta for enterprise clients. Using the Opus 4.7 model, this tool has already assisted in patching over 2,100 corporate vulnerabilities. Coalition partners like Cisco have also open-sourced tools such as the Foundry Security Spec to help global defenders build AI-assisted evaluation systems.
Recommendations for Organizations
Security teams should not rely solely on traditional patch cycles. Given the velocity of AI-driven vulnerability discovery, organizations are urged to:
- Enforce strict default configurations and principle of least privilege across all systems
- Mandate multi-factor authentication (MFA) on all critical access points
- Utilize advanced behavioral analytics to reduce mean time to detect (MTTD) post-breach activity
- Increase capacity for rapid vulnerability triage and emergency patching workflows
- Monitor threat intelligence feeds for newly disclosed CVEs matching your software stack
Project Glasswing represents a historic proof of concept for AI-driven security — and an urgent call for the industry to modernize its vulnerability management infrastructure before the AI discovery gap becomes a systemic threat.
Source: Cyber Security News, May 23, 2026