A cybersecurity startup managed to uncover weaknesses in Anthropic's internal systems using AI agents. The company says its core AI technology and customer conversations remained secure, but the findings raise bigger questions about how AI companies protect sensitive information.