Anthropic’s Glasswing project is proving to be a huge success. This project has made its cybersecurity technology Claude Mythos available on a preliminary basis to some companies and the result is that 10,000 high or critical severity vulnerabilities have already been detected. This makes it clear that the power of Claude Mythos was not hot air, as it is indeed finding a large number of security holes, faster and better than humans.
Around 50 companies and organizations have had privileged access to Claude Mythos, a cybersecurity tool that uses artificial intelligence (AI) to identify computer vulnerabilities.
Of the 10,000 vulnerabilities detected, more than 6,000 have been considered high or critical severity vulnerabilities. A more thorough analysis has revealed that there are more than a thousand truly valid positive cases.
Many detected vulnerabilities had been hidden for years and even decades, without being detected, at least apparently. Mythos has proven that very old security issues exist in operating systems, browsers and other core components.
Thanks to Mythos, the challenge now is not so much finding vulnerabilities as finding resources to fix them all. As a result of this initiative, Symfony has reported that Mythos found 19 vulnerabilities that have reportedly already been fixed. Drupal has also announced a major security patch. Obviously, to remain secure, updates must be carried out, something that not all web content management systems (CMS) do with due haste.
Given the increased ease of finding security holes with the new latest generation AI tools, it is absolutely essential for organizations and companies to also act with great speed when publishing fixes for those holes. Otherwise, companies such as banks could face catastrophic consequences.
Technologies such as Mythos Preview and GPT-5.5-Cyber have not been released to the public because they are really good at detecting security flaws. First, they are being used preventively to avoid misuse.
* Original article written in Spanish, translated with AI and reviewed in English by Jorge Mediavilla.
