AI company Anthropic cast its new Claude Mythos Preview model as the start of a cybersecurity reckoning, arguing that artificial intelligence is now becoming powerful enough to transform both digital defense and digital attack at the same time.
The company paired that warning with the launch of Project Glasswing, a tightly controlled initiative that gives select partners early access to the model for defensive security work.
Anthropic said Mythos Preview marks a "watershed moment for security," saying the model has shown unusual strength in computer security tasks and can identify and exploit software weaknesses at a level that could reshape the balance between defenders and attackers.
The company said that is precisely why it does not plan to make the model generally available right now. Instead, Anthropic is funneling access through Project Glasswing, which includes major partners such as Amazon Web Services, Apple, Broadcom, Cisco, CrowdStrike, Google, JPMorganChase, Microsoft, Nvidia, and Palo Alto Networks, along with broader access for about 40 additional organizations responsible for critical software infrastructure.
The effort is meant to strengthen defensive cybersecurity operations at a moment when governments and companies are increasingly worried about AI-powered hacking. Anthropic says the stakes are high because Mythos Preview has already uncovered thousands of vulnerabilities in foundational software, including major operating systems and web browsers.
Anthropic said the model found security problems "in every major operating system and web browser," and Anthropic's own materials say the project will focus on tasks such as local vulnerability detection, black-box testing of binaries, endpoint security, and penetration testing.
The company is also putting real money behind the initiative, saying it commit $100 million in model usage credits for Project Glasswing and related participants. It also announced $4 million in donations to open-source security organizations, including support through the Linux Foundation and the Apache Software Foundation, a sign that it wants to shore up the often underfunded software projects that underpin much of the internet.
The bigger message, though, is not just that AI can help fix bugs. It is the same systems that help defenders that may soon help adversaries move much faster. Anthropic said its long-term goal is to safely deploy "Mythos-class models" at scale, but only after improving safeguards that can detect and block the most dangerous outputs. Reuters also noted that Anthropic has been in discussions with the U.S. government about the model's capabilities, underscoring how seriously the company is treating the national security angle.
That caution follows weeks of attention around Mythos after reports about a leaked draft page describing the model's advanced cybersecurity capabilities. The company has since said the leak stemmed from human error in its content management system, not from a software breach.
