Progress from our Frontier Red Team shares insights into AI model risk assessment

AI Impact Summary

Progress from our Frontier Red Team is sharing insights into the rapidly evolving capabilities of AI models, particularly concerning cybersecurity and biosecurity risks. The team’s research demonstrates that models like Claude are showing ‘early warning’ signs of increasing skills in areas like cybersecurity, including solving CTF challenges at a level approaching undergraduate proficiency, and even autonomously executing cyberattacks with the aid of specialized tools. While these advancements are concerning, the team emphasizes that current models still fall short of thresholds that would represent substantially elevated risks to national security, highlighting the need for continued monitoring and mitigation strategies.