Anthropic has officially unveiled Mythos Preview, a powerful new AI model with advanced cybersecurity capabilities, alongside a massive industry-wide initiative called Project Glasswing. In a rare move, the AI startup is partnering with its primary competitors—including Google, Microsoft, and Apple —to ensure that the next generation of AI tools serves as a shield for digital infrastructure rather than a weapon for hackers.
The Rise of “Accidental” Cyber Capabilities
The development of Mythos Preview highlights a significant trend in artificial intelligence: emergent capabilities. Anthropic CEO Dario Amodei noted that the model was not specifically designed for hacking; rather, it was trained to be exceptionally proficient at coding.
As a direct byproduct of its coding prowess, the model has gained the ability to perform complex cybersecurity tasks, such as:
– Vulnerability discovery: Identifying weaknesses in software.
– Exploit development: Creating “proofs of concept” to demonstrate how a system can be breached.
– Penetration testing: Simulating attacks to test defenses.
– Binary analysis: Evaluating software without needing access to its original source code.
This creates a high-stakes “cat-and-mouse” game. While these tools can help defenders find and patch bugs faster, they also provide bad actors with the ability to automate sophisticated attacks that were previously too difficult or expensive to execute.
Project Glasswing: A Collaborative Defense
Recognizing the risk, Anthropic has formed Project Glasswing, a consortium designed to prepare the global tech ecosystem for a world of highly capable AI. The group includes more than 40 organizations, ranging from tech giants like Nvidia and Amazon Web Services to critical infrastructure and financial institutions.
The strategy behind this collaboration is two-fold:
1. Mitigating Vulnerabilities: By giving foundational platform developers (like Microsoft and Google) private access to Mythos Preview, they can use the model to find vulnerabilities in their own systems and patch them before the model is released to the general public.
2. Standardizing Defense: The group aims to redefine security paradigms. As Logan Graham, Anthropic’s frontier red team lead, explained, current security assumptions may break within the next 6 to 24 months as these capabilities become widely available.
“We’ve seen Mythos Preview accomplish things that a senior security researcher would be able to accomplish,” says Graham. “Done not carefully, this could be a meaningfully accelerant for attackers.”
Finding the “Unfindable” Bugs
The potential benefits of this collaboration are already being realized. Anthropic reports that Mythos Preview has already uncovered thousands of critical vulnerabilities, including certain bugs that had remained undetected in highly scrutinized code for decades.
For companies like Microsoft, the goal is to use AI to scale defense. Microsoft’s Global CISO, Igor Tsyganskiy, noted that as cybersecurity moves beyond purely human capacity, AI offers an “unprecedented” opportunity to reduce risk at a global scale.
The Path Ahead
Project Glasswing is currently in its early stages. For the initiative to succeed, it must move beyond a small circle of tech companies and evolve into a broader global standard for how AI-driven security is managed. The central challenge remains: identifying the critical questions of AI safety and finding answers before the technology outpaces our ability to defend against it.
Conclusion
By bringing competitors together under Project Glasswing, Anthropic is attempting to turn a potential security crisis into a coordinated defense effort. The goal is to ensure that as AI becomes more capable of breaking software, it becomes equally capable of protecting it.
