In a significant breakthrough for AI safety evaluation, the Mythos AI model (Preview) has become the first to successfully complete both cyber ranges established by the AI Security Institute (AISI). This achievement highlights the rapid evolution of testing frameworks designed to assess the offensive and defensive capabilities of next-generation artificial intelligence.
The AISI cyber ranges are specifically engineered to quantify and understand advanced cyberattack capabilities within AI systems. These rigorous benchmarks simulate complex, real-world hacking environments that demand sophisticated reasoning and vulnerability exploitation skills. While the highly anticipated GPT-5.5 model was only able to solve one of these intricate ranges during its evaluation, Mythos successfully cleared both, setting a new industry standard for high-capability models.
This demonstration underscores the critical importance of robust evaluation metrics as AI models continue to expand their technical boundaries. By pushing existing benchmarks, Mythos provides researchers and regulators with deeper insights into the potential risks and operational limits of advanced AI systems in cybersecurity contexts.
