The paper introduces AUTOATTACKER, a system designed to automate "hands-on-keyboard" cyber-attacks using large language models (LLMs). The authors address the challenges of automating complex cyber-attacks, such as tracking the execution environment, generating precise commands, and handling the large input space of attack commands. AUTOATTACKER consists of four main components: a summarizer, a planner, a navigator, and an experience manager. These components work together to generate and execute attack commands, leveraging the capabilities of LLMs like GPT-4, which demonstrate remarkable performance in automated post-breach attacks. The system is evaluated on a benchmark of 14 different attack tasks, covering various stages of the cyber kill chain, and shows high success rates, especially with GPT-4. The paper also discusses the limitations of prior works and the security issues of LLMs, emphasizing the need for robust methods to automate cyber-attacks while ensuring ethical and secure practices.The paper introduces AUTOATTACKER, a system designed to automate "hands-on-keyboard" cyber-attacks using large language models (LLMs). The authors address the challenges of automating complex cyber-attacks, such as tracking the execution environment, generating precise commands, and handling the large input space of attack commands. AUTOATTACKER consists of four main components: a summarizer, a planner, a navigator, and an experience manager. These components work together to generate and execute attack commands, leveraging the capabilities of LLMs like GPT-4, which demonstrate remarkable performance in automated post-breach attacks. The system is evaluated on a benchmark of 14 different attack tasks, covering various stages of the cyber kill chain, and shows high success rates, especially with GPT-4. The paper also discusses the limitations of prior works and the security issues of LLMs, emphasizing the need for robust methods to automate cyber-attacks while ensuring ethical and secure practices.