How Anthropic Learned Mythos Was Too Dangerous for the Wild

Anthropic has determined that its new AI model, Mythos, poses significant security risks and will restrict its release. According to a Bloomberg report, internal experts warned that Mythos could hack foundational computing systems, with researcher Nicholas Carlini stress-testing the model and finding its capabilities staggering. The company cited potential uses for espionage, theft, or sabotage as reasons for withholding the model from general availability.

The findings indicate Mythos represents a substantial capability leap in understanding and potentially exploiting system vulnerabilities. Carlini's assessment highlights the model's advanced proficiency with low-level computing infrastructure. Anthropic's decision follows its pattern of cautious deployment, though it continues to provide access to select vetted enterprises for specific use cases.

This development underscores ongoing tensions in AI development between advancing capabilities and managing associated risks. It reflects broader industry challenges in balancing innovation with safety protocols, particularly as models gain more sophisticated and potentially hazardous functionalities. The restricted release strategy aims to mitigate immediate threats while allowing controlled evaluation.