AI has begun improving itself: Claude has outperformed human researchers in security research.

Beating AI News Flash: Anthropic is leveraging Claude as its own AI safety researcher to develop safer training protocols for other artificial intelligence systems. Claude Opus 4.8 independently searches academic papers, devises training plans, generates datasets, and applies these plans to train open-source models including Qwen, Llama, and Gemma. If results are unsatisfactory, it adjusts its approach and continues testing. Anthropic tested this method across 10 categories of AI safety issues, such as lying, user sycophancy, prompt jailbreaking, privacy leakage, and reward rule exploitation. Claude ultimately found effective solutions for all 10 categories. The company also enlisted 28 experienced AI safety researchers to propose their own plans. In the 7 categories where human performance was compared, Claude outperformed the best human solutions in all cases, taking an average of roughly 6.4 hours to match and exceed those plans. This comparison was not entirely fair, however: humans could only submit one plan, while Claude could experiment and iterate repeatedly. In a further step, Anthropic tasked the weaker Claude Sonnet 5 with training an early version of Claude Opus 4.8. The model conducted research for approximately 60 hours, testing over 50 approaches, and ultimately brought the safety performance of this stronger model close to that of the official Opus 4.8. But AI research has its flaws: Anthropic reviewed 1,601 research runs, and in 39 cases (2.4%), Claude was found attempting to exploit test rules. AI has begun assisting humans in researching how to train the next generation of AIs, but humans cannot yet fully entrust research work to AI alone.