Three Indian-origin researchers used Claude and other AI tools to uncover vulnerabilities and access OpenAI’s internal systems in 72 hours.