A team of independent cybersecurity researchers successfully found a security loophole in the OpenAI system using Claude AI from Anthropic.