Anthropic Broadens AI-Driven Security Initiative To Shield Global Infrastructure From Emerging Cyber Threats

Anthropic announced on June 2, 2026, the expansion of Project Glasswing to approximately 150 organizations across more than 15 countries to identify and remediate critical software vulnerabilities. The initiative utilizes the Claude Mythos Preview model to scan open-source and proprietary codebases for high-severity security flaws that could be exploited by malicious actors.

This expansion represents a transition of the program from a targeted pilot into a broad international security framework designed to protect vital digital infrastructure. Because a single major breach in these interconnected systems could potentially impact over 100 million people and compromise national security, the project seeks to establish a proactive defense before offensive AI tools become more widespread. As automated cyberattacks grow in complexity, Anthropic argues that AI-driven defensive scanning is no longer a luxury but a fundamental necessity for maintaining global software integrity.

Geographic and Industrial Scope of the Expansion

The growth of Project Glasswing marks a significant escalation in Anthropic’s efforts to secure the global supply chain, moving from an initial cohort of 50 partners in April to 150 organizations today. This tripling of participants reflects a growing demand for automated security tools capable of keeping pace with rapid software development cycles. By including more than 15 countries in this phase, the initiative establishes a diversified defense network that spans various legal and technical jurisdictions.

A notable aspect of this expansion is the inclusion of South Korean firms and government entities, according to reports from Korea JoongAng Daily. The participation of these organizations suggests a strategic move to secure one of the world’s most technologically advanced regions against sophisticated regional and global threats. By integrating government agencies alongside private firms, Project Glasswing facilitates a public-private partnership model that is essential for defending national digital assets.

The strategic importance of including 15+ countries lies in the creation of a unified global defense posture. Software vulnerabilities do not respect national borders, and a flaw discovered in one region often exists in codebases used globally. By broadening the geographic footprint, Anthropic can gather a more comprehensive dataset of vulnerabilities across different coding standards and cultural development practices, strengthening the model’s overall efficacy.

The expansion also targets specific critical sectors, including power, water, healthcare, communications, and hardware. These sectors represent the foundational pillars of modern society, where a digital failure can have immediate and catastrophic physical consequences. Vulnerabilities in the power and water sectors, for instance, could lead to service interruptions that endanger public safety and health on a massive scale.

The inclusion of the hardware and communications sectors introduces unique risks compared to general software applications. Hardware vulnerabilities often involve firmware or microcode that is significantly more difficult to patch than standard application software, sometimes requiring physical access or specialized tools. In the communications sector, flaws can compromise the privacy of millions of users or allow for the interception of sensitive government and corporate data, making these high-priority targets for AI-driven defense.

Quantifying the Threat: 10,000 High-Severity Flaws Identified

Since the launch of the project in April, participants using Claude Mythos Preview have discovered more than 10,000 high- and critical-severity security flaws. This figure highlights the sheer volume of latent risks present in modern codebases, many of which have likely persisted for years despite traditional security audits. The density of these flaws—averaging roughly 200 critical issues per organization—reveals a troubling reality regarding the current state of open-source and proprietary software security.

This high volume of findings suggests that existing security practices have struggled to keep up with the complexity of modern software. Many of these 10,000 flaws likely reside in legacy code or deeply nested dependencies that are often overlooked during standard reviews. The fact that these were classified as “high” or “critical” severity indicates they could lead to full system compromise, unauthorized data access, or total service disruption if left unaddressed.

The primary tool driving these discoveries is Claude Mythos Preview, a specialized AI model designed for deep codebase analysis. Unlike traditional Static Analysis Security Testing (SAST) tools, which often rely on predefined patterns and rule sets, AI scanning can understand the context and logic of code. This allows the model to identify complex logic flaws and “zero-day” style vulnerabilities that do not match known signatures of previous attacks.

AI-driven scanning offers a significant advantage in terms of speed and depth compared to traditional methods. While a human security researcher might take weeks to audit a million lines of code, Claude Mythos Preview can perform similar tasks in a fraction of the time. Furthermore, the AI can correlate patterns across different parts of a codebase that a human or a standard SAST tool might fail to connect, providing a more holistic view of the application’s attack surface.

The discovery of 10,000 flaws in just a few months demonstrates that vulnerability identification is becoming less of a technical bottleneck. The challenge is shifting from finding the problems to managing the sheer volume of data produced by AI scanners. Organizations must now find ways to prioritize these 10,000 issues based on their potential impact and the likelihood of exploitation in a real-world environment.

The “Mythos-Class” Warning: A 6-to-12 Month Window

Anthropic has issued a stark warning regarding the future of the AI landscape, predicting that other companies will likely develop “Mythos-class” models within the next 6 to 12 months. This timeframe suggests an impending shift in the accessibility of high-tier AI capabilities, potentially ending the current period where such tools are restricted to a few controlled environments. The company expressed concern that these models could be released without the rigorous safety safeguards that Anthropic currently employs.

The risk of “Mythos-class” models being released without safeguards is that they could be repurposed for offensive cyber operations. In such a scenario, malicious actors could use the same logic-understanding capabilities to find and exploit vulnerabilities in target systems. This would lead to a world where cyberattacks occur much more frequently and in forms that are difficult for traditional defense systems to predict or block.

The democratization of these powerful tools poses a significant threat to small-to-medium enterprises (SMEs) that may not have access to programs like Project Glasswing. While large organizations and government entities are building AI-driven defenses, smaller firms often lack the resources to implement similar protections. If offensive AI tools become widely available while defensive tools remain restricted, the security gap between large and small organizations could widen dangerously.

Anthropic’s warning emphasizes the urgency for cyberdefenders to adapt their strategies immediately. The 6-to-12 month window represents a critical period for organizations to harden their infrastructure and automate their security pipelines. If the defensive community fails to keep pace, the arrival of unrestricted Mythos-class models could lead to a significant increase in successful data breaches and infrastructure disruptions.

The company maintains that it is imperative for the industry to maintain a “defense-forward” posture. This involves not only identifying flaws but also ensuring that the tools used to find them are handled responsibly. By sounding this alarm, Anthropic is calling for a broader industry discussion on the ethical release of high-capability AI models and the necessity of built-in safety protocols to prevent misuse in cyber warfare.

Operational Challenges: Solving the Remediation Bottleneck

As AI-driven discovery tools become more efficient, the primary bottleneck in cybersecurity is shifting from finding vulnerabilities to fixing them. Anthropic reports that while Claude Mythos Preview can identify thousands of flaws rapidly, security teams still face the labor-intensive tasks of verifying findings, disclosing issues, and developing patches. This human-centric phase of the process is currently the slowest link in the security chain.

Security teams are now tasked with managing an unprecedented influx of data. Each of the 10,000 identified flaws requires a human expert to confirm that the finding is not a false positive and to assess the specific risk it poses to their unique environment. This verification process is essential to ensure that limited development resources are not wasted on non-critical issues, but it also creates a significant drag on the overall speed of remediation.

The challenge extends to the patching and deployment phases, where the danger of “giant backlogs” becomes a reality. According to Anthropic, if the rate of discovery continues to outpace the rate of patching, organizations will find themselves with a massive list of known vulnerabilities that they cannot address quickly enough. This “remediation debt” provides a roadmap for attackers, who may attempt to exploit these known but unpatched flaws before the organization can secure them.

To address this human resource gap, organizations must find ways to scale their security teams or further automate the remediation process itself. This might include using AI to generate draft patches or to automate the testing of those patches before they are deployed to production systems. However, even with these tools, the final responsibility for software integrity remains with human developers, who must oversee the deployment to ensure no new bugs are introduced.

The current operational reality highlights the need for a fundamental change in how security teams function. Moving from a manual, reactive model to an automated, proactive one requires not just new tools but also new workflows. Organizations must integrate AI discovery directly into their continuous integration and continuous deployment (CI/CD) pipelines to ensure that vulnerabilities are addressed as soon as they are found, rather than waiting for a periodic security audit.

Strategic Tools and Future Accessibility

To assist in these efforts, Anthropic has introduced “Claude Security,” a tool that utilizes the Claude Opus 4.8 model for advanced codebase scanning and patching support. This tool is designed to bridge the gap between discovery and remediation by providing developers with actionable insights and suggested fixes for the vulnerabilities found. By leveraging the more powerful Opus 4.8 architecture, Anthropic aims to provide a higher level of accuracy and more sophisticated patching recommendations.

Currently, Anthropic provides these vulnerability discovery tools primarily to trusted security teams and approved partners upon request. This “trusted partner” model allows the company to monitor the use of its models and ensure they are being used for defensive purposes. It also provides a controlled environment to refine the models’ performance and safety features before a wider release is considered.

This controlled approach contrasts with the eventual goal of general availability. By limiting access initially, Anthropic can prevent the tools from being used by actors who might seek to exploit the vulnerabilities found rather than fix them. This strategy reflects the company’s commitment to safety and its recognition of the dual-use nature of high-capability AI models in the cybersecurity domain.

Organizations seeking access to these tools must meet specific security standards and demonstrate a commitment to responsible disclosure. This requirement ensures that the entities using Project Glasswing have the necessary infrastructure and ethical frameworks to handle the sensitive information generated by the scans. Such standards are necessary to prevent the accidental leakage of vulnerability data, which could be just as damaging as a direct attack.

The long-term strategy involves expanding access as the models and their associated safeguards become more robust. Anthropic is working toward a future where these defensive tools are a standard part of the software development lifecycle for all organizations. However, the transition from a restricted pilot to a general-market product will depend on the successful mitigation of the risks identified during the current phase of Project Glasswing.

Closing

Anthropic’s roadmap for Project Glasswing includes the eventual release of Mythos-class models to all customers, but only after robust safeguards are finalized and proven effective. This cautious approach is intended to ensure that the defensive advantages of AI are not overshadowed by the risks of exploitation. The company continues to monitor the performance of Claude Mythos Preview across its 150 partner organizations to gather the data necessary for these safety improvements.

The long-term outlook for the cybersecurity landscape suggests a permanent arms race between AI-driven defense and AI-driven exploitation. As defensive tools become more sophisticated, attackers will undoubtedly seek to use AI to find new ways around them. Success in this environment will require constant innovation and a global commitment to sharing threat intelligence. Project Glasswing serves as a foundational step in building the collaborative, AI-powered infrastructure needed to protect the digital world in the years to come.