Monday AI Radar #22

The UK AI Safety Institute's evaluation of Anthropic's Mythos model reveals a significant leap in AI-powered cybersecurity threats. Unlike previous systems that could identify vulnerabilities, Mythos demonstrates the ability to autonomously conduct complete cyberattacks consisting of numerous discrete steps—finding vulnerabilities, exploiting them, and chaining attacks together without human intervention. This represents what experts call a 'Mythos moment' for cybersecurity, where AI capabilities fundamentally change the threat landscape.

Security experts warn that bio-AI capabilities likely follow close behind cyber advancements, creating parallel threats in multiple domains. The evaluation suggests we have approximately 1-2 years before these capabilities become widespread, with the timeline heavily dependent on mundane but critical factors like rapid patch deployment. Meanwhile, AI safety discussions continue with debates around 'alignment-by-default'—the theory that LLMs trained on human text have inherent value alignment—though experts caution this doesn't eliminate alignment challenges.

The newsletter also highlights concerns about 'mundane misalignment' in current models that could scale to catastrophic levels if unaddressed. While alignment remains unsolved, the concrete nature of these challenges—from autonomous cyberattacks to value alignment—represents progress in identifying and addressing AI risks that were largely theoretical just 5-10 years ago.