BT explainer: Anthropic’s Claude Mythos preview is here to reshape cybersecurity—here’s how

Anthropic, renowned for its AI innovations, has unveiled a preview of its latest model, Claude Mythos, designed to detect and exploit software vulnerabilities. Unlike previous models, this powerful AI has not been released to the general public due to security concerns regarding its capabilities. Dubbed Claude Mythos, this model represents a significant leap in reasoning, coding, and autonomous planning, specifically tailored for cybersecurity applications. Its name, inspired by the Ancient Greek term for the narrative systems that shape civilizations, reflects the model's ambitious purpose. What sets Claude Mythos apart? According to Anthropic, it stands as their most advanced model to date. While earlier iterations like Claude 3.5 and Opus 4.6 demonstrated their usefulness as supportive tools, Mythos operates as an independent security researcher. In practical scenarios, it has successfully identified critical vulnerabilities in major operating systems and web browsers. In a recent blog post, Anthropic revealed that Mythos Preview has uncovered thousands of previously unknown zero-day vulnerabilities, some of which had eluded software developers for decades. Notable discoveries include a 27-year-old flaw in OpenBSD, which could permit attackers to crash machines remotely, and a hidden 16-year-old bug within FFmpeg video software. Additionally, Mythos has shown the ability to combine vulnerabilities in the Linux kernel, granting unauthorized users full control over systems. The AI model achieved impressive results, scoring 94% on the SWE-benchmark testing and outperforming its predecessor, Claude Opus 4.6, with a score of 83.1% compared to its 66.6%. With its robust capabilities, Claude Mythos is being integrated into a new initiative called Project Glasswing, which focuses on cybersecurity. This project is accessible to a select group of about 40 organizations, including tech giants like Google, Microsoft, and Apple, as well as security leaders such as CrowdStrike and Palo Alto Networks. Financial institutions like Goldman Sachs and JPMorgan Chase are also exploring the model's potential to anticipate cyber threats. Anthropic has committed up to $100 million in usage credits and an additional $4 million in donations to open-source security initiatives, reflecting its dedication to enhancing cybersecurity. Currently, the general public does not have access to the Mythos model. Reports suggest that its powerful capabilities necessitate the implementation of protective measures before broader release. Anthropic plans to refine and test new cybersecurity protocols, potentially introducing them with the next Claude Opus model before expanding Mythos's availability. This strategic approach aims to bolster security across various industries, enabling companies to proactively identify vulnerabilities and fortify their systems against cyber threats.