Anthropic's AI model accessed real systems during tests, revealing cybersecurity flaws. The incident highlights the need for stronger safeguards as AI capabilities grow, prompting calls for regulatory action.