“Too powerful to be released to the public,” but already used by over 40 companies to test and strengthen the security of critical software. This is Claude Mythos Preview, the new artificial intelligence model from Anthropic, which has identified thousands of vulnerabilities in widespread systems, including flaws in all major operating systems and browsers, raising the alarm about a new era of AI-related threats.
STRATEGY AND PARTNERS OF THE GLASSWING CONSORTIUM
To contain the risks, Anthropic decided not to make the technology public and to collaborate with a consortium of companies known as Project Glasswing. Partners include Apple, Amazon, Microsoft, Google, Cisco, and Broadcom, as well as the Linux Foundation, committed to open-source software.
Anthropic has allocated up to 100 million dollars in Claude usage credits for the project, allowing participants to share the results obtained with Mythos and to strengthen critical infrastructures and open-source software.
MODEL CAPABILITIES AND RESULTS
As explained by the New York Times, Claude Mythos Preview was not specifically trained for cybersecurity, but its advanced programming and reasoning capabilities allow it to identify vulnerabilities previously difficult to detect even by human experts.
According to Logan Graham, head of Anthropic’s team testing models for dangerous capabilities, “the model identifies vulnerabilities understandable to security experts, but it has also discovered flaws and created exploits so sophisticated that they have eluded researchers and all automated tools for decades.” Among the most significant vulnerabilities, Anthropic states, are a 27-year-old bug in OpenBSD and a defect in video software tested five million times without ever being detected.
A GLOBAL SECURITY ALERT
For Jared Kaplan, Chief Science Officer of Anthropic, “the goal is both to raise awareness and to give responsible parties an advantage in protecting infrastructures and code, both open-source and private.” The project’s name, Glasswing, refers to the glass butterfly, with transparent wings that camouflage, a metaphor for software vulnerabilities hidden in complex systems but present for years. The initiative highlights how AI models are rapidly transforming the ability to detect and fix flaws, creating both defensive opportunities and risks for infrastructures and sensitive data.
WHAT THE EXPERTS SAY
According to Elia Zaitsev, CTO of CrowdStrike, a cybersecurity company, “what used to take months now happens in minutes thanks to AI,” and Claude Mythos Preview shows how defenders can act on a large scale, while noting that adversaries will inevitably try to exploit the same capabilities.
Mike Krieger of Anthropic added that the model is used as a defensive tool, “almost like arming them in advance,” allowing the open-source community and cybersecurity specialists to intervene before vulnerabilities can be exploited.
ANTHROPIC’S BACKGROUND
Anthropic, founded in 2021 by former OpenAI executives, has gained global attention for balancing the development of advanced AI systems with risk management. The company, according to the NYT, stated that its projected annual revenue for 2026 exceeds 30 billion dollars, more than triple the previous 9 billion, thanks to Claude’s popularity as a programming tool. The company has already had clashes with the Pentagon, which considers it a supply chain risk, and discussions are ongoing, with some restrictions suspended by a federal judge.
