Anthropic has revealed that its Project Glasswing security initiative has surfaced more than 10,000 high- or critical-severity software vulnerabilities since the effort launched roughly a month ago, putting fresh pressure on developers to patch faster. The disclosure, made on Friday, underscores how quickly AI-assisted vulnerability discovery is reshaping the economics of cyber defense.
What Project Glasswing is
Project Glasswing gives a small group of about 50 partners early access to Claude Mythos Preview, a frontier model built to autonomously hunt for flaws in widely used software before attackers can weaponize them. Anthropic frames the program as a defensive push aimed at hardening the most "systemically important" software that underpins global infrastructure.
Of the flaws flagged so far, 6,202 were initially classed as high- or critical-severity across more than 1,000 open-source projects. Follow-up analysis confirmed 1,726 as valid true positives, with as many as 1,094 rated high- or critical-severity. The work has so far led to 97 fixes landing upstream and 88 advisories being issued.
A critical WolfSSL flaw and a blocked $1.5M fraud
Among the most serious discoveries is a critical weakness in WolfSSL, tracked as CVE-2026-5194 with a CVSS score of 9.1, that could let an attacker forge certificates and impersonate a legitimate service. Anthropic also said the model proved useful beyond bug hunting: one Glasswing partner bank reportedly used it to detect and stop a fraudulent $1.5 million wire transfer after an intruder compromised a customer's email and placed spoofed phone calls.
Why defenders are being told to move faster
The findings arrive as vendors ship more patches than ever, a trend that mirrors other AI-driven security efforts. Anthropic is urging organizations to compress patch testing and deployment timelines, harden default configurations, enforce multi-factor authentication, and keep thorough logs for detection and response. The company also launched a Cyber Verification Program letting vetted security professionals use its models without guardrails for legitimate tasks such as penetration testing and red teaming, echoing OpenAI's comparable Daybreak program for GPT-5.5-Cyber.
An asymmetric advantage, for now
Models like Claude Mythos Preview remain unreleased to the public over fears that adequate safeguards do not yet exist to prevent large-scale misuse. Anthropic argues that Glasswing hands the most important defenders an "asymmetric advantage," but warns that comparable capabilities could soon be widely available. The move continues a busy stretch for the company, which recently expanded its research leadership. It also lands amid a wave of attacks on software ecosystems, from the Megalodon GitHub Actions campaign to Microsoft's takedown of a malware-signing service.
Reporting based on coverage from The Hacker News and Anthropic.