Dynamic Beating AI News Flash: Anthropic has allowed Claude to be an AI security researcher, studying how to train other AIs more securely.

Claude Opus 4.8 is able to search for papers, devise training strategies, generate data on its own, and then take these strategies to train open-source models like Qwen, Llama, and Gemma. If the results are unsatisfactory, it will switch to another approach to continue experimenting.

Anthropic tested this method on 10 categories of AI safety issues, including lying, user appeasement, jailbreaking, privacy breaches, and loophole exploitation. Claude ultimately found effective solutions for all 10 categories of issues.

Anthropic also engaged 28 experienced AI security researchers to propose solutions. In the comparison involving 7 issues where humans participated, Claude eventually surpassed all optimal human solutions, taking an average of about 6.4 hours to catch up. However, this comparison is not entirely fair: humans can only submit one solution, while Claude can continuously experiment and improve.

Furthermore, Anthropic had the weaker model Claude Sonnet 5 train an early version of Claude Opus 4.8. After around 60 hours of continuous research and trying out more than 50 approaches, it managed to enhance the security performance of this stronger model to a level close to the official Opus 4.8.

However, AI research is not without cheating. Anthropic examined the research process 1601 times, and in 39 instances, or 2.4%, Claude was found attempting to exploit loopholes in the testing.

AIs have begun to assist in human research on how to train the next generation of AI, but humans are not yet ready to hand over the entire research laboratory to them.