Project Glasswing: How Tech Giants Are Securing Critical Software for the AI Era

The AI race has a shadow nobody wants to talk about. While headlines obsess over which model writes better code or passes harder exams, a quieter, more dangerous race is underway. The same AI that builds software faster than any human also learns to break it faster. And the defenders? They've been losing ground for years.

That changed, or at least, started to change, on April 7, 2026. Anthropic announced Project Glasswing, an unprecedented alliance that brings together Amazon, Apple, Broadcom, Cisco, CrowdStrike, the Linux Foundation, Microsoft, Palo Alto Networks, and roughly 40 other organizations. Their mission? Use advanced AI to find and fix security flaws in the world's most critical software before attackers do.

It's a bold move. And honestly, it's about time.

---

What Is Project Glasswing? A Defensive Alliance for the AI Age

At its core, Project Glasswing is a defensive cybersecurity initiative. Anthropic is granting exclusive access to Claude Mythos Preview, an unreleased frontier AI model, to a carefully selected group of partners responsible for critical software infrastructure. These partners use the model to scan first-party and open-source systems for vulnerabilities, then share their findings with the broader industry.

Think of it like a neighborhood watch, but instead of looking for suspicious people, the neighbors are some of the world's largest tech companies, and they're armed with an AI that spots weaknesses human eyes (and traditional tools) have missed for decades.

The name itself tells a story. Glasswing butterflies have transparent wings that make them nearly invisible, a metaphor for software vulnerabilities, which are "relatively invisible" until someone knows where to look. That's the premise: use AI to make the invisible visible.

There's a catch, though. Anthropic isn't releasing Mythos Preview to the general public. The model's capabilities are powerful enough that in the wrong hands, it could be weaponized for cyberattacks. "We really do view this as a first step for giving a lot of cyber defenders a head start on a topic that will be increasingly important," Dianne Penn, Anthropic's head of research product management, explained.

---

Claude Mythos Preview, The Engine Behind the Mission

So what makes Mythos Preview so special? Here's the surprising part: it wasn't specifically trained for cybersecurity. It's a general-purpose frontier model with exceptional coding and reasoning capabilities, the kind of AI that Anthropic calls its "most powerful" yet.

And yet, those general capabilities have proven remarkably effective at identifying subtle security flaws. Over just a few weeks of testing, Mythos Preview identified thousands of zero-day vulnerabilities, many of them critical and hiding in widely-used software for years.

Why does this matter? Because it suggests something profound: general AI reasoning, not narrow, purpose-built security tools, may be our best defense against increasingly sophisticated threats. The same capabilities that make AI dangerous when misused also make it invaluable when aimed at protection.

---

The Invisible Threat: Why Software Security Looks Different in the AI Era

Here's a number that should keep you up at night: open-source software constitutes the vast majority of code in modern systems, including the very systems AI agents use to write new software.

But those open-source maintainers? They're often volunteers with shoestring budgets and no access to advanced security tools. Meanwhile, AI coding assistants are churning out more code faster than ever, and AI-powered attacks are evolving at machine speed.

The BSIMM16 report, the industry's most comprehensive snapshot of software security practices, confirms what many already suspected: AI coding is the new reality, and it's further destabilizing software supply chain security. Organizations in 2026 need to secure their AI coding while simultaneously defending against AI-enabled attacks.

In other words, we're fighting a two-front war: protect the AI we're building, and protect ourselves from the AI others are building against us.

---

AIBOMs and SBOMs, Why You Can't Secure What You Can't See

There's an old cybersecurity adage: you can't protect what you can't see. In the AI era, that means knowing exactly what's in your software supply chain.

Enter the AI Bill of Materials, or AIBOM. Think of it like a nutrition label for AI systems, a complete inventory of every model, plug-in, training dataset, and third-party dependency in use. The OWASP AIBOM Generator is an open-source tool designed to make this visibility practical, generating AIBOMs (also called AI SBOMs or ML-BOMs) to enhance supply chain transparency and security.

Project Glasswing doesn't directly address AIBOMs, but it tackles the same underlying problem: invisibility is the attacker's greatest advantage. Whether it's a decades-old bug in OpenBSD or a poisoned dependency in an AI model, the first step to fixing it is finding it.

---

The NIST Cyber AI Profile, A Framework for the New Normal

Frameworks matter. They give organizations a shared language and a structured way to think about risk. And in 2026, NIST stepped up with its preliminary draft of the Cyber AI Profile, guidance that maps AI-specific cybersecurity considerations to the six CSF functions: Govern, Identify, Protect, Detect, Respond, and Recover.

The message is clear: AI security isn't a separate discipline. It's integrated into existing cybersecurity and risk frameworks, with trust, identity, and cryptographic assurance underpinning AI governance.

Projects like Glasswing are the offensive (well, defensive) counterpart to this framework work, real-world implementation of the principles frameworks describe.

---

Real-World Impact: What Project Glasswing Has Already Uncovered

Theory is nice. Results are better. And Project Glasswing has already delivered some jaw-dropping finds.

A 27-year-old bug in OpenBSD. OpenBSD is an operating system known primarily for its security focus. Yet for nearly three decades, a vulnerability lurked in its code, undetected by human reviewers and automated tools alike, until Mythos Preview flagged it.

A 16-year-old vulnerability in FFmpeg. FFmpeg is a widely-used video software library. Automated testing tools had run the affected code line five million times. Five. Million. Times. They never caught the flaw. Mythos Preview did.

These aren't edge cases. They're evidence that our existing security tooling is fundamentally limited — and that AI can see patterns and relationships that traditional static analysis and fuzzing tools miss.

Anthropic is putting real resources behind this: up to $100 million in usage credits for the project, plus $4 million in direct donations to open-source security organizations. All vulnerabilities discovered have been patched, with maintainers contacted and fixes deployed.

---

What This Means for Security Teams, Developers, and Organizations

So you're not Anthropic. You don't have access to Mythos Preview. What do you actually do with this information?

For CISOs and security leaders: Securing the AI supply chain needs to be a top priority. That means understanding where your AI models come from, what dependencies they pull in, and how they're being used. The CISO priority list for 2026 includes continuous visibility into the entire AI supply chain, because determining why a model makes a wrong decision becomes critical when that error might come from a threat actor's interference.

For developers: AI coding assistants aren't going away. But the BSIMM16 findings suggest organizations need to secure their AI coding practices while defending against AI-enabled attacks. This means security reviews that account for AI-generated code, provenance tracking, and controls that don't just find issues but prove they've been addressed.

For open-source maintainers: Project Glasswing represents something unprecedented: access to a new generation of AI models that can analyze codebases at a scale and depth previously impossible. If you maintain critical open-source infrastructure, keep an eye on how findings from this initiative are being shared.

For everyone: The "durable advantage" concept is worth sitting with. Anthropic's framing, that Project Glasswing is "an important step toward giving defenders a durable advantage in the coming AI-driven era of cybersecurity", acknowledges that the advantage won't come from a single tool or project. It will come from building defensive capabilities that keep pace with offensive ones.

---

The Road Ahead, Limitations and Open Questions

Let's be honest: Project Glasswing is a starting point, not a solution. Anthropic itself says so: "Project Glasswing is a starting point. No single organization can solve these cybersecurity problems alone."

What remains to be seen:

• Scalability. Can this model of controlled, partner-based access scale to protect the entire software ecosystem?
• Broader access. Will similar capabilities eventually become available to smaller organizations and individual developers?
• Regulatory alignment. How will initiatives like this interface with frameworks like the NIST Cyber AI Profile and emerging AI governance requirements?
• Offense vs. defense balance. The industry predicts similar AI capabilities will become more widespread within two years. Will defenders keep their head start?

---

A Durable Advantage Starts with Collective Action

The AI era is rewriting the rules of software security. The code that runs our world, from banking systems to power grids to the apps on your phone, is more complex, more interconnected, and more vulnerable than ever.

Project Glasswing is a signal. It says that the companies building frontier AI recognize the responsibility that comes with it. It says that collaboration, even among fierce competitors, is possible when the stakes are high enough.

But a signal isn't a solution. The durable advantage Anthropic talks about won't come from any single model or any single company. It will come from an entire ecosystem, AI developers, software companies, security researchers, open-source maintainers, and governments, working together to make the invisible visible, and the vulnerable secure.

Want to stay informed about AI security developments like Project Glasswing? Subscribe to our newsletter for weekly insights on the intersection of artificial intelligence, cybersecurity, and critical infrastructure.