Anthropic has developed an advanced artificial intelligence model that it has decided not to release publicly. The model, called Claude Mythos Preview, has identified "thousands of vulnerabilities" in software, some of which have been hidden for decades, according to the company itself. To address potential risks, Anthropic has launched the Glasswing project, a collaboration with 12 leading companies that allows these organizations to use Mythos Preview to fix flaws before they can be exploited by malicious actors or similar models. Nicholas Carlini, a scientist at Anthropic, stated in a presentation that Mythos Preview has detected more vulnerabilities in two weeks than he has in his entire career. Vulnerabilities are inadvertent flaws in software that can be exploited to access or manipulate systems. Traditionally, discovering these errors required extensive knowledge and time, but Mythos Preview has proven to be much more efficient. Among the companies participating in Glasswing are tech giants like Apple, Amazon Web Services, Google, Microsoft, Nvidia, and Crowdstrike. Anthropic has provided access to Mythos Preview to these companies and 40 others that manage critical infrastructures, in addition to holding discussions with the U.S. government. Originally, Mythos Preview was intended to be a model focused on code generation, following the success of Claude Sonnet 4.5 in September. However, according to Dario Amodei, CEO of Anthropic, its extraordinary capabilities in cybersecurity soon became evident. Among the vulnerabilities already addressed by Mythos Preview are a 27-year-old flaw in OpenBSD, a 16-year-old vulnerability in the FFmpeg library used in video applications, and several flaws in the Linux kernel that allowed privilege escalation to full system control. The name Mythos comes from ancient Greek and, according to Anthropic, refers to the "system of stories with which civilizations interpreted the world." While some consider the warning about risks to be a marketing strategy, there are precedents: in 2019, OpenAI delayed the release of GPT-2, deeming it "too dangerous," and Dario Amodei, now CEO of Anthropic, was leading research at OpenAI at that time. Whereas there were previous concerns that AI models could be used to create misinformation or spam, the risk is now more tangible: the exploitation of software vulnerabilities. Mythos Preview not only identifies flaws but can also generate the necessary exploits to take advantage of them. Additionally, external experts have verified its effectiveness. Beyond its cybersecurity capabilities, Mythos is a versatile model that improves upon its predecessors. Anthropic has published a 244-page manual detailing the differences and advancements compared to earlier models like Sonnet and Opus. Recently, Anthropic also made headlines for its dispute with the U.S. Department of Defense, which considered it a "supply chain risk," preventing it from contracting with the government. However, the company won the litigation, maintaining its conditions of not allowing the use of its AI for mass surveillance or for autonomous weapons without oversight.