“A recent research paper co-authored with OpenAI, Google DeepMind, and Anthropic found that human attackers can break 100% of AI model defenses within 10 to 30 attempts.”