"Superhuman" Go AIs still have trouble defending against these simple exploits

Researchers at MIT and FAR AI have discovered vulnerabilities in top-level AI Go algorithms that allow humans to exploit gaps and defeat the AI using unorthodox strategies. Despite efforts to improve the algorithms' defenses against such attacks, including fine-tuning models and iterative training, the results have shown limited success. The study highlights the challenge of creating truly robust and unexploitable AI systems, even in controlled environments like board games. The findings suggest that while AI algorithms can excel in average performance, they remain vulnerable to simple exploits in worst-case scenarios. The research emphasizes the importance of addressing vulnerabilities in AI systems to prevent embarrassing failures when deployed to the public. Despite the difficulties encountered in defending against attacks in Go, the researchers remain optimistic about the potential to enhance AI robustness through continued research and training against a variety of attacks.