WayToClawEarn
High impactAnthropic

Anthropic Project Glasswing update: AI discovers over 10,000 high-risk vulnerabilities in one month

Anthropic released the first month results of Project Glasswing: Claude Mythos Preview found more than 10,000 high-risk vulnerabilities, and 90.6% of open source software was confirmed to be real and effective. AI security auditing moves from proof-of-concept to production.

WayToClawEarn EditorialPublished May 23, 2026Updated Aug 8, 2026

Editorial review of public sources · AI-assisted drafting. How we work · Original source

Core conclusion

On May 22, 2026, Anthropic released a phased results update for Project Glasswing. Only one month after its launch, the Claude Mythos Preview model has discovered more than 10,000 high-risk or critical vulnerabilities together with about 50 partners, covering the Internet's most critical infrastructure software. Meanwhile, Mythos Preview's scans of more than 1,000 open source projects found a cumulative 23,019 vulnerabilities, of which 6,202 were assessed as high or critical.

Core changes: AI discovers vulnerabilities far faster than humans can patch them—"verification, disclosure, and repair" have become the new bottleneck. This is both a challenge and an opportunity for the software security ecosystem.

Key Points

  • Event Time: 2026-05-22 (officially released by Anthropic)
  • Scope of Impact: Global software security ecosystem, open source maintainers, enterprise security teams
  • Core Signal: The practical capabilities of AI in the field of cybersecurity have far exceeded most people’s expectations, and security practitioners must re-evaluate threat models and response speeds

Background and trigger events

In April 2026, Anthropic officially launched Project Glasswing - a joint effort with approximately 50 partners to leverage the power of the Claude Mythos Preview model to harden the security of the world's most critical infrastructure software before AI can be exploited for malicious purposes. Anthropic has previously emphasized that Mythos Preview has reached new heights in network security capabilities.

This May 22 update is Glasswing’s first official results report since its launch, and for the first time makes public actual usage data from partners and external third-party verification results.

Key Impact (by Dimension)

DimensionsChangeWhat it means to usRecommended actions
Vulnerability discovery efficiencySome partners have increased vulnerability discovery speed by more than 10 timesAI-assisted code auditing is no longer a proof-of-concept, but can be put into productionIntegrate AI security scanning tools into the development pipeline as soon as possible
Open source software securityMythos scanned 1,000+ projects and found 23,019 vulnerabilities, 90.6% were effective true positivesThe open source libraries you rely on may have unpatched vulnerabilitiesCheck the vulnerability reports that the project depends on to shorten the patch cycle
Patch cycle pressureEach high-risk vulnerability takes an average of 2 weeks to patchThe speed of discovery is faster than the speed of repair, and the attack window period is lengthenedEstablish an automated patch CI/CD pipeline to shorten the test deployment cycle
Enterprise Security PracticesCloudflare discovers 2,000 vulnerabilities with Mythos, with false positive rate better than human testersThe role of human security auditors is shifting from "finding vulnerabilities" to "classifying and remediating"Redefining the workflow and staffing of security teams
Financial security practiceAssisted a cooperative bank to prevent a fraudulent transfer of $1.5 millionAI is not only used for code audits, but can also directly participate in real-time security defenseIntroducing AI security assistance in sensitive business links

Adaptation suggestions

Implications for content production teams

An intriguing detail from Anthropic's Glasswing project: Mozilla found 271 vulnerabilities in Firefox 150 using Mythos Preview, 10 times more than Claude Opus 4.6 found in Firefox 148. This means that model capabilities are experiencing leaps and bounds in the security field. For teams that rely on AI Agents for content automation, you can expect the same jump in capabilities to occur with content accuracy, security, and consistency.

Executable points

  • Review your tool chain: Assess the security of the AI technology stack you use in the following dimensions: code auditing, dependency management, permission control
  • Updated Security Policy: A Mythos-level model means vulnerability discovery is no longer a scarce resource—the real scarcity is the ability to fix quickly. Establish automated quality gates and patch pipelines
  • Focus on the Dilemma of Open Source Maintainers: Reports indicate that some open source maintainers have explicitly requested Anthropic to slow down disclosures (because they need more time to design patches). If you rely on critical open source projects, consider making a donation or contributing labor

Example: Real vulnerability case of Mythos Preview

Mythos Preview discovered a high-risk vulnerability in wolfSSL, an open source encryption library used by billions of devices: AI successfully built attack code that exploited this vulnerability. An attacker can forge certificates and allow users to access seemingly normal banking or email service websites, but are actually controlled by the attacker. The vulnerability has since been patched (CVE-2026-5194), but this incident clearly demonstrates that AI is capable of combat-grade exploits.

Next action

View source →

Disclaimer: this site shares educational insights only, for inspiration and reference. No outcome guarantee; external execution and decisions are your own responsibility.