AI Starts Finding Vulnerabilities on Its Own! Anthropic Project Glasswing Expands to 15 Countries, Claude Mythos Discovers Over 10,000 0-Day Vulnerabilities in a Month

At 11 PM, a security engineer at a networking equipment company in Hsinchu was still staring at a scan report on his computer. The static analysis tool he was using had been running for a whole day, flagging dozens of potential issues, most of which were false positives. He rubbed his eyes, thinking to himself: "If only there was something that could directly help me identify the real vulnerabilities, that would be great."

That "something" is what Anthropic claims to have developed. On June 2, 2026, Anthropic announced the significant expansion of its security project, Project Glasswing, adding approximately 150 organizations across 15 countries. The project utilizes a yet-to-be-publicly-released model, Claude Mythos, to proactively discover vulnerabilities in critical infrastructure code. Not only can it find vulnerabilities, but it can also write functional attack programs to verify them. AI has started finding vulnerabilities on its own, and this is transitioning from research to reality.

Event Background

Project Glasswing is not a sudden development. Its core is Claude Mythos, Anthropic's most powerful model designed specifically for finding security vulnerabilities in code. Due to the dual nature of this capability (finding vulnerabilities can also mean exploiting them), Anthropic has chosen not to publicly release Mythos, making it available only to defensive alliance members.

Before this expansion, the early Glasswing alliance had already included over 50 tech organizations, such as Microsoft, Apple, Google, and Cloudflare. The results were impressive: within the first month, Mythos Preview automatically discovered over 10,000 high-risk and critical 0-day vulnerabilities. Mozilla used it to fix 271 vulnerabilities in Firefox 150, a tenfold increase in findings compared to using Claude Opus. It even found a critical vulnerability in the wolfSSL encryption library (CVE-2026-5194) and successfully created an attack to verify it. The UK AI Security Institute noted that Mythos was the first model to fully solve their "multi-step network attack simulation."

Key Points This Time

• Significant Scale Expansion: On June 2, Glasswing expanded to approximately 150 new organizations across 15 countries, including Australia, Canada, France, Germany, Italy, Switzerland, the Netherlands, Spain, Belgium, Sweden, India, Japan, New Zealand, and South Korea.
• Impressive Partner List: New partners include Okta, Samsung, SK Hynix, SK Telecom, as well as government and defense-level organizations like NATO and the EU's cybersecurity agency, ENISA.
• Focus on Critical Infrastructure: The target is concentrated on the power, water, healthcare, communication, and hardware industries. Anthropic states that these partners have in common the potential for severe consequences if successfully attacked, with most cases potentially affecting over 100 million people.
• Remarkable but Controlled Capability: Mythos can find tens of thousands of vulnerabilities and even automatically generate usable exploits within weeks. Due to its danger, Anthropic does not publicly release it, restricting its use to the defensive alliance.

Market Impact Analysis

For Taiwanese Users: Generally, people won't directly use Mythos, but they will indirectly benefit. The vulnerabilities in the browsers, operating systems, and encrypted communications they use daily are being identified and fixed by AI ahead of time, reducing the number of potential attack vectors. For example, Firefox fixing 271 vulnerabilities at once is a tangible benefit. However, this shouldn't lead to complacency—basic security practices like updating software, using two-factor authentication, and being cautious with links are still essential. Tools like Whoscall can provide an additional layer of protection against scams and misinformation.

For Enterprises: This is the most critical segment for Taiwanese businesses. With a large number of hardware, semiconductor, and networking companies, the participation of SK Hynix and Samsung in the alliance signals that competitors in the same industry are already utilizing "AI-powered vulnerability discovery" for defense. For Taiwanese companies, this is a clear signal: the traditional pace of annual penetration testing may no longer be sufficient. While not every company may be able to join such a top-tier alliance in the short term, they should at least begin evaluating the integration of AI-assisted security checks into their processes. For guidance on how to assess and secure AI agents before deployment, refer to AI Agent 上線前,你一定要做的評測與安全把關.

For Developers: This is both good news and pressure. The good news is that AI's vulnerability discovery capability will gradually be integrated into general development tools, potentially providing deeper security suggestions within tools like GitHub Copilot, Cursor, or Claude in the future. The pressure is that the era of "code that just runs" is passing—when AI can automatically find vulnerabilities in code, writing secure code will become a basic requirement rather than an extra credit item. To understand how the AI coding landscape is changing, see Claude Code 領跑、微軟 Google 急追:2026 年中 AI 寫程式戰局解析.

Future Development Trends

First, the era of AI vs. AI in cybersecurity has begun. Defensive parties use AI to find and fix vulnerabilities, while attackers will use AI to find and exploit them, escalating the speed and scale of cybersecurity battles.

Second, the debate over whether powerful models should be publicly released will intensify. Anthropic's decision to keep Mythos restricted to alliance members for defensive use will be discussed and potentially emulated by others, raising questions about the governance of "controlled models."

Third, critical infrastructure will be prioritized for protection, but small and medium-sized enterprises may be left behind. Top-tier alliances focus resources on targets like power, healthcare, and communication, which could affect over 100 million people if attacked. General small and medium-sized enterprises will have to wait until these capabilities are integrated into more accessible tools.

TheAI學院 Summary and Commentary

Honestly, what impresses me most about this news isn't the "10,000 vulnerabilities found" but Anthropic's decision to keep Mythos restricted. A model powerful enough to automatically generate attack programs, if made publicly available, could have unforeseeable consequences. The restraint shown by not releasing it publicly is, in itself, a statement of intent.

AI not only writes code but now also finds code vulnerabilities on its own—cybersecurity's offense and defense have officially entered the "machine vs. machine" era.

For Taiwanese readers, the concrete advice is: if you're an IT or security officer, you should already have "AI-assisted security checks" on your evaluation list for this year. Don't wait until your industry peers have adopted it to follow. If you're a developer, from today on, consider "security" as the default when writing code, not something to be addressed just before launch.