Claude Mythos Preview: the new AI model from Anthropic that redefines the limits of cybersecurity Anthropic has officially unveiled Claude Mythos Preview, its latest advancement in artificial intelligence, with unusual caution: access will be restricted to selected technology partners only. This decision, while puzzling to many, reflects the magnitude of the model's capabilities and the associated risks. An unprecedented leap in cyber intelligence. On February 24, 2026, Anthropic engineers tested Claude Mythos Preview for the first time and were astonished by its performance. The model demonstrated an unprecedented ability to autonomously identify and exploit zero-day vulnerabilities in major operating systems and browsers, far surpassing its predecessors. Implications for global security. This advancement, while valuable for digital defense, poses a significant risk if it falls into the wrong hands. Anthropic highlighted in its blog that Mythos was able to discover a vulnerability in OpenBSD that had gone undetected for 27 years, despite the system's reputation for security. Examples like this illustrate the potential and danger of the tool. Unrivaled performance. According to the technical report published by Anthropic, Claude Mythos Preview has outperformed leading models such as GPT 5.4, Gemini 3.1 Pro, and Claude Opus 4.6 in most benchmarks, particularly excelling in mathematical tests like USAMO, where it approaches perfection. Dramatic reduction in hallucinations. The model exhibits a significantly lower error rate than its predecessors and can recognize when it lacks sufficient information, minimizing incorrect responses due to overconfidence. Subtle errors, a new challenge. However, the report warns that when Mythos fails in complex tasks, its errors are so sophisticated that they may go unnoticed even by experts, necessitating thorough review of its responses. Project Glasswing: restricted access. Due to its power, Claude Mythos Preview will only be available through the defensive Glasswing program, reserved for companies like AWS, Apple, Broadcom, Cisco, CrowdStrike, Google, JPMorgan Chase, the Linux Foundation, Microsoft, NVIDIA, and Palo Alto Networks. These organizations will be responsible for using the model to detect and correct vulnerabilities before they can be exploited. A starting point, not an endpoint. Anthropic emphasizes that Mythos Preview does not represent the pinnacle of AI capabilities in cybersecurity, but rather a new beginning. The company anticipates even greater advancements in the coming years. Source: xataka.com