This research investigates the use of autonomous systems driven by Large Language Models (LLMs) for Assumed Breach penetration testing in enterprise networks. The authors developed a novel prototype capable of compromising accounts within a real-life Active Directory testbed. Their evaluation highlights the prototype's strengths and limitations in simulating attacks, using a realistic environment to capture complex network behaviors. The study concludes that autonomous LLMs show promise for democratizing access to penetration testing. The prototype's code and analysis are publicly released to foster further research in LLM-driven cybersecurity automation.