Project Glasswing: the initiative aiming to shield the world’s most critical software in the AI era

Project Glasswing has arrived in a big way, and it’s no exaggeration to say it could change the cybersecurity game for good. Anthropic just announced an initiative that brings together some of the biggest names in global technology — including AWS, Apple, Microsoft, Google, Cisco, NVIDIA, CrowdStrike, Broadcom, JPMorganChase, Palo Alto Networks, and the Linux Foundation — all united by a common goal: protecting the planet’s most critical systems before it’s too late.

The trigger for all of this was a discovery that’s both impressive and, honestly, a little scary. A new artificial intelligence model called Claude Mythos Preview, not yet available to the public, demonstrated a capability few expected to see this soon: finding and exploiting software vulnerabilities with a level of precision and autonomy that surpasses most human security experts.

We’re not talking about simple bugs or obvious mistakes here. The model found thousands of high-severity vulnerabilities, including flaws in all major operating systems and web browsers. Issues that had been hiding for years or even decades, surviving human reviews and millions of automated tests. The good news is that many have already been patched. But the message is clear: if an AI can do this defensively, what happens when those same capabilities fall into the wrong hands?

That’s exactly the question that motivated Project Glasswing, and in the following sections you’ll understand how this initiative works, what it’s already uncovered, and why the time to act is now. 🔐

What is Project Glasswing and why it matters so much

Project Glasswing isn’t just another bug bounty program or a corporate PR initiative. It represents a real shift in how the tech industry views cybersecurity in the age of artificial intelligence. The core idea is simple but powerful: if AI models are becoming capable of finding flaws that humans can’t detect, then those same models need to be put to work for defense — in a coordinated, transparent way with shared responsibility among the most influential companies in the sector.

Anthropic took the lead in this conversation after realizing that Claude Mythos Preview had reached a level of technical capability that went far beyond what was expected from a general-purpose language model. When internal researchers began testing the model in code analysis scenarios, the results exceeded expectations: the model didn’t just identify problematic patterns — it could trace complete exploitation paths, understand the usage context of the code, and suggest precise fixes, all autonomously and at speeds far exceeding those of a conventional human team.

Instead of keeping the technology for internal use or simply publishing a technical report, Anthropic chose to open the dialogue with strategic partners already operating at the most critical layers of global digital infrastructure. AWS, Apple, Microsoft, Google, Cisco, NVIDIA, and CrowdStrike aren’t just lending their names to the project. They’re contributing access to real-world environments, telemetry data, proprietary architectures, and specialized teams, creating a collaborative defense network that has never existed before at this scale.

Beyond the launch partners, Anthropic extended access to a group of more than 40 additional organizations that build or maintain critical software infrastructure, so they can use the model to scan and protect both proprietary and open-source systems. The company is also committing up to $100 million in usage credits for Mythos Preview across these initiatives, along with $4 million in direct donations to open-source security organizations. That’s Project Glasswing in action. 🛡️

Cybersecurity in the age of artificial intelligence

The software we all depend on every single day — running banking systems, storing medical records, connecting logistics networks, keeping power grids running, and so much more — has always contained bugs. Many are minor, but some represent serious security flaws that, if discovered, can allow cyberattackers to hijack systems, disrupt operations, or steal data.

The consequences of cyberattacks on corporate networks, healthcare systems, energy infrastructure, transportation hubs, and government agencies are already well documented. On the global stage, state-sponsored attacks have threatened to compromise the infrastructure that underpins both civilian life and military readiness. Even smaller-scale attacks, like those targeting individual hospitals or schools, can still cause substantial economic damage, expose sensitive data, and even put lives at risk. The global financial costs of cybercrime are difficult to estimate precisely, but they may run around $500 billion per year.

Historically, many software flaws went unnoticed for years because finding and exploiting them required expertise that only a handful of security specialists possessed. With the latest frontier AI models, the cost, effort, and level of expertise required to find and exploit vulnerabilities have dropped dramatically. Over the past year, AI models have become increasingly effective at reading and reasoning about code, showing a remarkable ability to identify vulnerabilities and figure out ways to exploit them.

Claude Mythos Preview demonstrates a leap in these cyber capabilities. Ten years after the first DARPA Cyber Grand Challenge, frontier AI models are becoming competitive with the best humans at finding and exploiting vulnerabilities. Without the right safeguards, these powerful cyber capabilities could be used to exploit the many existing flaws in the world’s most important software, making cyberattacks of all kinds far more frequent and destructive.

Despite the serious risks, there’s reason for optimism: the same capabilities that make AI models dangerous in the wrong hands make them invaluable for finding and fixing flaws in critical software, as well as producing new programs with far fewer security bugs. 💡

Claude Mythos Preview: the AI that finds what humans can’t see

To understand the weight of what Claude Mythos Preview represents, you need to know what it has already accomplished. In recent weeks, Anthropic used the model to identify thousands of zero-day vulnerabilities — meaning flaws that were previously unknown to the software developers themselves — many of them critical, across all major operating systems and web browsers, plus a variety of other important software.

The discovered vulnerabilities weren’t typos or misconfigurations. They were deep structural flaws, the kind that require very specific technical knowledge to even be recognized as a problem. The model was able to identify nearly all of these vulnerabilities and develop many related exploits completely autonomously, without any human guidance. Three examples illustrate the caliber of these discoveries:

• OpenBSD: Mythos Preview found a 27-year-old vulnerability in OpenBSD, an operating system with a reputation for being one of the most secure in the world, used to run firewalls and other critical infrastructure. The flaw allowed an attacker to remotely crash any machine running the operating system simply by connecting to it.
• FFmpeg: A 16-year-old vulnerability was discovered in FFmpeg, a library used by countless software applications to encode and decode video. The flaw was in a line of code that automated testing tools had hit five million times without ever catching the problem.
• Linux Kernel: The model found and autonomously chained together multiple vulnerabilities in the Linux kernel — the software that runs most of the world’s servers — allowing an attacker to escalate from regular user access to full machine control.

All of the vulnerabilities above were reported to the respective software maintainers and have already been patched. For many other vulnerabilities, Anthropic is providing a cryptographic hash of the details and will disclose specifics after fixes have been implemented.

What sets Claude Mythos apart from traditional static analysis tools is its ability for contextual reasoning. Conventional tools look for known patterns and only find what has already been cataloged. Mythos Preview, on the other hand, can infer system behaviors from how different parts of the code interact with each other, simulating usage scenarios that no developer considered during the original writing. This represents a significant technical breakthrough, putting artificial intelligence in a position to discover the unknown rather than merely confirming what’s already known.

Results on evaluation benchmarks like CyberGym reinforce the substantial gap between Mythos Preview and Anthropic’s next-best model, Claude Opus 4.6. The model also achieved the highest scores ever recorded on a variety of software coding tasks, including SWE-bench Verified, Pro, and Multilingual, as well as Terminal-Bench 2.0 and other relevant benchmarks. 🔍

What partners are saying about the project

Several partner organizations have already had access to Claude Mythos Preview for a few weeks, and their statements reinforce both the urgency and the potential of the initiative.

Cisco highlighted that AI capabilities have crossed a threshold that fundamentally changes the urgency needed to protect critical infrastructure against cyber threats, and that legacy methods of system hardening are no longer sufficient.

AWS emphasized that its teams analyze over 400 trillion network flows per day in search of threats and that AI is central to their ability to defend at scale. The company had already been testing Claude Mythos Preview in its own security operations and applying it to critical codebases.

Microsoft stressed the unprecedented opportunity to use AI responsibly to improve security and reduce risk at scale. When tested against CTI-REALM, Microsoft’s open-source security benchmark, Claude Mythos Preview showed substantial improvements compared to previous models.

CrowdStrike warned that the window between a vulnerability being discovered and being exploited by an adversary has collapsed: what used to take months now happens in minutes with AI.

The Linux Foundation noted that open-source maintainers, whose software underpins much of the world’s critical infrastructure, have historically been left to handle security on their own. Project Glasswing offers a path to change that equation, turning AI-augmented security into an accessible tool for all maintainers.

JPMorganChase emphasized that advancing cybersecurity and the resilience of the financial system is central to its mission, and that the project offers a unique opportunity to evaluate next-generation AI tools for cyber defense in critical infrastructure.

Google highlighted its ongoing investment in AI-powered security tools, such as Big Sleep and CodeMender, to find and fix critical software flaws, reinforcing its active participation in the initiative.

Palo Alto Networks warned that the model represents both a radical shift in finding previously hidden vulnerabilities and a dangerous signal that attackers may soon find even more zero-day vulnerabilities and develop exploits faster than ever before. The recommendation is clear: all organizations need to prepare for AI-assisted attackers. 📢

What changes in practice for global cybersecurity

One of the most concrete changes Project Glasswing promises to deliver is faster response times for critical software vulnerabilities. Today, the cycle of discovery, responsible disclosure, and patching can take weeks or even months, depending on the complexity of the issue and the organization involved. With models like Claude Mythos integrated into continuous analysis pipelines, that cycle could be reduced dramatically. The model doesn’t sleep, doesn’t have selective attention bias, and doesn’t get overwhelmed by code volume. It analyzes, prioritizes, and reports consistently, which is already a massive advantage over the current model.

Beyond speed, there’s also the coverage factor. Most organizations don’t have the resources to maintain security teams that can continuously review every software component they use, especially when they depend on open-source libraries maintained by volunteers. Project Glasswing addresses exactly this blind spot. By bringing together partners with a presence across different layers of digital infrastructure, the project creates a more complete view of systemic risks, identifying which components are most widely used, which are least monitored, and where the impact of a flaw would be most devastating.

In practice, Project Glasswing partners will receive access to Claude Mythos Preview to find and fix vulnerabilities or weaknesses in their foundational systems — systems that represent a very large portion of the world’s shared cyber attack surface. The work should focus on tasks like local vulnerability detection, black-box binary testing, endpoint protection, and penetration testing.

Another important aspect is the impact on security culture within participating companies. When organizations the size of Apple, Google, and Microsoft publicly commit to an initiative like this, they’re also signaling internally that cybersecurity is a strategic priority — not just an operational cost. This tends to influence hiring decisions, budget allocation, and even how engineering teams are incentivized to treat security as part of the development process rather than a final step. 🌐

Tools we use daily

Translation

Text Inspection & Clipping

Social Media

Search & Research

Recruitment

Programming

Productivity & Organization

Operations

Marketing & Sales

Finance & Investment

Education

Design & Media

Data & Analytics

Content & Writing

Automation

Investment, access, and the initiative’s next steps

Anthropic’s financial commitment to Project Glasswing is significant. The $100 million in usage credits for the model will cover substantial use during this initial research phase. After that period, Claude Mythos Preview will be available to participants at $25 per million input tokens and $125 per million output tokens, with access via Claude API, Amazon Bedrock, Google Cloud Vertex AI, and Microsoft Foundry.

Beyond usage credits, Anthropic donated $2.5 million to the Alpha-Omega and OpenSSF projects through the Linux Foundation, and $1.5 million to the Apache Software Foundation, so that open-source software maintainers can respond to this shifting landscape. Maintainers interested in access can apply through the Claude for Open Source program.

Anthropic intends for this work to grow in scope and continue for many months. Partners will share information and best practices with one another to the extent possible. Within 90 days, Anthropic will publish a report on what it has learned, the vulnerabilities patched, and the improvements made that can be disclosed. The company will also collaborate with leading security organizations to produce a set of practical recommendations on how security practices should evolve in the AI era, potentially including:

• Vulnerability disclosure processes
• Software update processes
• Open-source and supply chain security
• Software development lifecycle and secure design practices
• Standards for regulated industries
• Triage scalability and automation
• Patch automation

Anthropic also shared that it has been in ongoing discussions with U.S. government officials about Claude Mythos Preview and its offensive and defensive cyber capabilities. The company acknowledges that protecting critical infrastructure is a top national security priority for democratic nations.

Regarding the model’s general availability, Anthropic does not plan to make Claude Mythos Preview accessible to the general public. The eventual goal is to allow users to deploy Mythos-class models safely and at scale, both for cybersecurity purposes and for the many other benefits that models this capable will bring. To get there, the company needs to advance the development of safeguards that detect and block the model’s most dangerous outputs. Anthropic plans to release new safeguards with a future Claude Opus model, allowing them to improve and refine those safeguards with a model that doesn’t carry the same level of risk as Mythos Preview. 📊

Why now is the right time

Artificial intelligence has reached a point where its offensive and defensive capabilities are evolving at the same speed, but defense mechanisms are still disorganized, fragmented, and reactive. Project Glasswing emerges as an attempt to shift that balance, putting the most advanced tools available in the hands of those who want to protect systems — not compromise them. The initiative acknowledges something the industry was slow to admit: no single company can handle the complexity and scale of the risks that arise when artificial intelligence begins operating at the same technical level as the best human specialists.

Claude Mythos Preview is living proof that this moment has already arrived. The software vulnerabilities it found weren’t hypothetical. They were real, they were in production, and they could have been exploited. The fact that they were discovered and fixed before causing harm is a win, but also a wake-up call. The next discovery might not come from a collaborative project with good intentions. It could come from someone who developed a similar capability in secret, with no commitment to responsible disclosure or the collective good.

As Anthropic itself put it, the work of defending the world’s cyber infrastructure could take years, but frontier AI capabilities will likely advance substantially in just a few months. For cyber defenders to stay ahead, action needs to happen now.

This is the context in which Project Glasswing takes on its deepest meaning. It’s not just a technical response to a technical problem. It’s a declaration that the global tech community can choose how this transition unfolds — and that the choice to act in a coordinated, open, and responsible manner has real, measurable, and urgent value. Cybersecurity has never been as dependent on collective decisions as it is right now, and initiatives like this show that at least part of the industry has gotten the message. Anthropic has invited other members of the AI industry to join the effort to establish industry standards, and in the medium term, an independent, third-party body — bringing together organizations from both the public and private sectors — could be the ideal home for continuing this kind of project at scale. ⚡