Project Glasswing and Claude Mythos: How Effective Is Anthropic’s AI at Finding Critical Vulnerabilities?

First‑Month Results: More Than 10,000 Vulnerabilities

Anthropic’s first update on the program reported unusually large‑scale results:

• More than 10,000 high‑ or critical‑severity vulnerabilities discovered collectively by Anthropic and roughly 50 partner organizations.
• Hundreds of serious vulnerabilities per partner discovered during roughly the first month of use.
• Several organizations reporting their bug‑finding rate increased by more than 10× after using the system.

One example cited by Anthropic is Cloudflare, which reported discovering about 2,000 bugs—including 400 high‑ or critical‑severity issues—across its core services using the model.

These figures suggest the system may allow security teams to analyze large codebases much faster than traditional manual auditing.

Benchmark Performance and Technical Capabilities

External evaluation signals also point to strong cybersecurity performance.

The UK AI Security Institute reportedly confirmed that Claude Mythos Preview solved 73% of expert‑level capture‑the‑flag cybersecurity challenges, a benchmark used to test real‑world vulnerability‑finding skills.

Anthropic and related analyses also describe several capabilities:

• Discovering previously unknown zero‑day vulnerabilities in operating systems and browsers.
• Chaining multiple vulnerabilities together to achieve privilege escalation or system compromise.
• Automatically generating working exploit code to demonstrate how flaws could be abused.

These abilities move beyond simple static analysis and into tasks traditionally performed by elite security researchers.

Examples of Vulnerabilities Found

While most findings remain undisclosed during coordinated patching, a few examples have been reported.

According to reports summarizing Anthropic’s claims, the model identified:

• A 27‑year‑old vulnerability in the OpenBSD kernel related to TCP SACK options.
• A 16‑year‑old bug in FFmpeg that had survived millions of automated tests without being detected.
• Chains of vulnerabilities in the Linux kernel that allowed privilege escalation from a normal user to full system control.

Anthropic has said it cannot disclose details for the vast majority of discoveries yet because over 99% of the vulnerabilities it found have not been publicly patched.

Which Companies Are Involved

Project Glasswing operates as a coalition of technology companies, infrastructure providers, and security organizations.

Reported launch participants include:

• Amazon Web Services (AWS)
• Apple
• Google
• Microsoft
• NVIDIA
• Cisco
• Broadcom
• CrowdStrike
• Palo Alto Networks
• JPMorgan Chase
• The Linux Foundation

Along with Anthropic itself and dozens of additional organizations, the program involves roughly 50 partners in total.

These partners run the model against their own software and critical open‑source infrastructure to identify vulnerabilities that can then be patched.

Why the Model Is Not Publicly Released

Anthropic has emphasized that Mythos is intentionally restricted because the same capabilities useful for defense could also enable large‑scale cyberattacks.

If widely available, a system capable of automatically discovering and exploiting zero‑day vulnerabilities could dramatically lower the barrier for offensive cyber operations.

The controlled partnership model is intended to fix as many vulnerabilities as possible before similar capabilities become broadly available.

Criticism and Skepticism From Security Experts

Despite the impressive headline numbers, several points of skepticism remain within the cybersecurity community.

1. Limited independent verification
Most of the large‑scale results—such as the 10,000 vulnerabilities discovered—come from Anthropic’s own reporting or summaries repeating those claims. Because the vulnerabilities are mostly undisclosed, outside researchers cannot yet evaluate them directly.

2. Possible overlap with existing tools
Some researchers argue that similar vulnerability‑finding capabilities may already be achievable using combinations of existing public models and tools, suggesting Mythos may be an acceleration rather than a fundamentally new capability.

3. Difficulty measuring real‑world validation rates
Although benchmark results such as the 73% capture‑the‑flag success rate exist, clear statistics on how many AI‑identified bugs were confirmed and patched in real production systems have not yet been publicly documented in detail.

The Bottom Line

If Anthropic’s reported numbers hold up under broader verification, Project Glasswing represents one of the largest AI‑assisted vulnerability discovery efforts ever conducted. In its first month alone, the program reportedly uncovered more than 10,000 high‑severity security flaws across critical software systems.

However, because most vulnerabilities remain undisclosed during coordinated patching, the security community still lacks comprehensive independent validation of these results.

What is clear is that AI systems are becoming powerful tools for software security research—capable of scanning complex codebases, identifying subtle vulnerabilities, and even generating working exploits. Whether Mythos represents a unique breakthrough or simply the leading edge of a broader trend in AI‑driven security analysis is a question the industry will likely answer over the next few years.