Open-weights AI with Mythos-level cyber skills likely in 24 months

The post argues there’s an 85% chance an open-weights AI model—or system—capable of Mythos-level cybersecurity will emerge in the next 24 months. The reasoning is based on the global distribution of open-weights AI labs, including in the US, France, UAE, South Korea, and Canada, all of which currently trail the frontier by no more than 24 months. Given the geopolitical fragmentation and competing national interests, a coordinated international ban on open-weights releases appears implausible before such models emerge.

Even if a ban were attempted, the post highlights the near-impossibility of preventing exfiltration of model weights from closed-weights labs, which would also possess Mythos-level capabilities. The author speculates that today’s open-weights frontier—potentially represented by models like Kimi K3—could be fine-tuned and optimized to achieve Mythos-level performance in cybersecurity or other domains with sufficient compute and infrastructure. This raises urgent questions about societal preparedness for such capabilities.