AI Unleashed: Anthropic’s Secret Weapon Finds Thousands of Hidden Zero-Days in Everyday Tech

In a move that could redraw the battle lines of cybersecurity, Anthropic has quietly launched Project Glasswing-a bold collaboration harnessing the power of artificial intelligence to hunt down the world’s most elusive software vulnerabilities. At the heart of this initiative is Claude Mythos Preview, an unreleased AI model that’s already rattling the nerves of both hackers and defenders by uncovering thousands of zero-day flaws across virtually every major operating system and browser on the planet.

Fast Facts

• Anthropic’s Project Glasswing brings together industry giants like AWS, Apple, Google, and Microsoft to leverage AI for cybersecurity defense.
• The new AI model, Claude Mythos Preview, has autonomously discovered thousands of high-severity zero-day vulnerabilities-including in OpenBSD, FFmpeg, and the Linux kernel.
• Unlike previous models, Mythos’s skills emerged organically, not through targeted security training.
• Access to Mythos is tightly controlled and limited to vetted organizations managing critical infrastructure.
• Anthropic has committed $100 million in usage credits and additional funding to support open-source security efforts.

The AI That Outsmarts the Experts

With cybercrime inflicting an estimated $500 billion in annual damages worldwide, the stakes for digital defense have never been higher. Project Glasswing’s centerpiece, Claude Mythos Preview, represents a seismic leap in AI-driven security. Unlike its predecessors, Mythos wasn’t specifically trained to hunt vulnerabilities. Instead, its advanced reasoning, coding, and autonomy have spontaneously produced skills that leave even seasoned human experts in the dust.

Take OpenBSD, for example-a system revered for its security. Mythos unearthed a 27-year-old flaw that could have allowed attackers to remotely crash any machine running the OS, simply by connecting to it. In the case of FFmpeg, a ubiquitous video library, the AI spotted a subtle bug that had evaded detection through five million automated tests over 16 years. Most alarmingly, Mythos independently chained together multiple Linux kernel vulnerabilities, demonstrating how an attacker could escalate from a regular user to total system control-all with minimal human prompting.

These discoveries aren’t just theoretical. Maintainers have already patched the highlighted flaws, underscoring the immediate impact-and value-of the initiative. On the CyberGym benchmark, Mythos reproduced 83.1% of known vulnerabilities, a significant jump from the 66.6% achieved by Anthropic’s earlier Claude Opus 4.6 model.

But with great power comes great risk. Anthropic is acutely aware that the same AI capabilities that empower defenders could just as easily be turned against them by malicious actors. As a result, Mythos remains under lock and key, accessible only to a select group of over 40 organizations responsible for safeguarding the world’s most critical software infrastructure.

Backing the project is a war chest of $100 million in AI usage credits and millions more to support open-source security, a tacit acknowledgment that no single entity can face the coming storm alone. Project Glasswing signals a new era-one where alliances between tech titans, open-source maintainers, and government agencies may be the last line of defense against ever-evolving AI-powered threats.

The Road Ahead

Project Glasswing’s launch is more than a technical milestone-it’s a warning shot across the bow of complacency in cybersecurity. As AI’s offensive capabilities accelerate, so too must our collective response. The digital world just got a lot more complicated-and a lot more dangerous. Whether the defenders can stay one step ahead remains to be seen, but with tools like Claude Mythos Preview, the balance of power may finally be shifting.

WIKICROOK

• Zero: A zero-day vulnerability is a hidden security flaw unknown to the software maker, with no fix available, making it highly valuable and dangerous to attackers.
• Operating System (OS): An Operating System (OS) is core software that manages computer hardware and enables applications to run, ensuring smooth and secure operation.
• Kernel: The kernel is the core of an operating system, managing hardware and software resources to ensure efficient and secure system operation.
• Automated Testing: Automated testing uses software tools to run security tests on code or systems, identifying vulnerabilities quickly and reducing the need for manual checks.
• Frontier Model: A frontier model is a highly advanced AI system, trained on large datasets, capable of sophisticated reasoning and analysis in cybersecurity contexts.