Project Glasswing: Anthropic Initial Update on AI-Driven Vulnerability Discovery

TL;DR

Anthropic has launched Project Glasswing, a collaborative initiative using the Claude Mythos Preview model to secure critical software infrastructure. In its first month, the project and its partners identified over 10,000 high- or critical-severity vulnerabilities, demonstrating that AI can now find bugs at a rate that far exceeds human capacity to verify and patch them.

Performance of Claude Mythos Preview

Claude Mythos Preview has demonstrated a significant leap in autonomous cybersecurity capabilities, enabling partners to find vulnerabilities at rates often exceeding ten times their previous capacity.

Partner and External Validation

  • Aggregate Impact: Approximately 50 partners have collectively found more than 10,000 high- or critical-severity vulnerabilities.
  • Cloudflare: Reported finding 2,000 bugs (400 high- or critical-severity) in critical-path systems, noting a false positive rate superior to human testers.
  • UK AI Security Institute: Confirmed Mythos Preview is the first model to solve both of its cyber ranges (multistep cyberattack simulations) end-to-end.
  • Mozilla: Identified and fixed 271 vulnerabilities in Firefox 150, a tenfold increase compared to the findings from Claude Opus 4.6 in Firefox 148.
  • Benchmarks: The model is the strongest performer on the ExploitBench and ExploitGym academic benchmarks and showed "unprecedented precision" on XBOW's web exploit benchmark.

Real-World Application

Beyond codebase scanning, Mythos Preview was used by a partner bank to detect and prevent a fraudulent $1.5 million wire transfer resulting from a compromised email account and spoofed phone calls.

Open-Source Software Scanning Results

Anthropic used Mythos Preview to scan over 1,000 open-source projects. The model identified an estimated 23,019 total vulnerabilities, including 6,202 estimated as high- or critical-severity.

Verification and Accuracy

To validate these findings, six independent security research firms and Anthropic assessed a subset of 1,752 high- or critical-rated vulnerabilities. The results showed:

  • True Positive Rate: 90.6% (1,587) were valid vulnerabilities.
  • Severity Confirmation: 62.4% (1,094) were confirmed as high- or critical-severity.

Case Study: wolfSSL

Mythos Preview detected a vulnerability in the wolfSSL cryptography library (CVE-2026-5194) that would have allowed attackers to forge certificates to host fake websites for banks or email providers. This vulnerability has since been patched.

The "Patching Bottleneck" and Systemic Risks

Project Glasswing has revealed that the primary constraint in cybersecurity is no longer the discovery of vulnerabilities, but the human capacity to triage, report, and deploy patches.

The Remediation Gap

While AI can find bugs almost instantly, the industry standard Coordinated Vulnerability Disclosure (CVD) policy typically involves a 90-day window for patching. This creates a dangerous interim period where vulnerabilities are known to the AI (and potentially attackers) but not yet patched by users.

Impact on Maintainers

Open-source maintainers are currently overloaded. Many have reported being capacity-constrained by a "deluge of low-quality, AI-generated bug reports," leading some to request that Anthropic slow down the rate of disclosures.

Adaptation Strategies for Cyberdefenders

As Mythos-class models become more widely available, Anthropic recommends that developers and network defenders shift their strategies to reduce exposure.

Recommendations for Developers

  • Shorten Patch Cycles: Accelerate the time between bug discovery and the availability of security fixes.
  • Improve Update UX: Make it easier for end users to install updates and be more persistent in notifying users running vulnerable software.

Recommendations for Network Defenders

  • Accelerate Deployment: Shorten the timeline for testing and deploying patches.
  • Implement Layered Controls: Rely on NIST and NCSC frameworks to harden default configurations, enforce multi-factor authentication (MFA), and maintain comprehensive logs.

New Defensive Tools and Programs

Anthropic is releasing several tools to help organizations leverage AI for defense without requiring the capabilities of the restricted Mythos Preview model.

Claude Security and Verification

  • Claude Security (Public Beta): Available for Claude Enterprise customers, this tool uses Claude Opus 4.7 to scan codebases and propose fixes. It has already been used to patch over 2,100 vulnerabilities in three weeks.
  • Cyber Verification Program: Allows security professionals to use models for legitimate research, penetration testing, and red-teaming without certain misuse safeguards.

Technical Resources for Security Teams

Qualifying customers can now request a suite of tools used by Project Glasswing partners, including:

  • Custom Skills: Instructions for repeated security tasks.
  • Scanning Harness: A system to map codebases, deploy scanning subagents, and write reports.
  • Threat Model Builder: A tool to prioritize model work by mapping potential attack targets.

Future Outlook and Model Release

Anthropic has not yet released Mythos-class models to the public because current safeguards are insufficient to prevent severe misuse. Project Glasswing serves as a proactive effort to harden critical software before such capabilities are broadly available.

Moving forward, Anthropic intends to expand Project Glasswing to more partners, including US and allied governments, and will release Mythos-class models only after developing significantly stronger safeguards.

Sources

Related