Strategic Warnings on AI Risk Progress and Insights from Frontier Red-Teaming

Frontier AI models are showing rapid capability improvements in dual-use domains such as cybersecurity and biology, sending 'early warning' signals of national security risk.

The progress in cybersecurity in particular stands out. Claude's CTF (Capture The Flag) problem-solving ability has rapidly improved from high-school level to undergraduate level over the past year.

The latest model, Claude 3.7 Sonnet, raised its solve rate on the Cybench benchmark from about 5% to about 33%. This improvement is observed in the following areas:

• Exploitation (pwn)
• Web applications (web)
• Cryptography (crypto)

However, it has not yet reached expert level. Reverse engineering of binary executables and reconnaissance and attacks in network environments remain challenging. In addition, physical constraints, specialized equipment, and human expertise continue to act as barriers that suppress actual risk.