According to @godofprompt on Twitter, a joint study by OpenAI, Anthropic, and Google systematically tested current AI safety defenses—such as prompting, training, and filtering models—against advanced and adaptive attacks, including gradient descent, reinforcement learning, random search, and human red-teamers (Source: arxiv.org/abs/2510.09023, @godofprompt). Despite previous claims of 0% failure rates, every major defense was bypassed with over 90% success, with human attackers achieving a 100% breach rate where automated attacks failed. The study exposes that most published AI defenses only withstand outdated, static benchmarks, failing to address real-world attack adaptability. These findings signal a critical vulnerability in commercial LLM applications, warning businesses that current AI security solutions provide a false sense of protection. The researchers stress that robust AI defense must survive both RL optimization and sophisticated human attacks, urging the industry to invest in dynamic and adaptive defense strategies.