ISCAP Proceedings: Abstract Presentation
REMAP - A Containerized Platform for Teaching Students to Author Agentic Reverse Engineering Tools
Michael Ham
Dakota State University
Abstract
The National Security Agency's National Centers of Academic Excellence in Cybersecurity (NCAE-C) program designates institutions that offer rigorous cybersecurity and AI curricula aligned with prescribed Knowledge Units (KUs); Software Reverse Engineering is a mandatory KU. As large language models (LLMs) increasingly enhance security workflows, the ability to develop and expand AI tools, rather than merely use them, has become an important skill for employers. Yet most classroom exposure to AI-assisted reverse engineering treats these tools as black boxes: students prompt a model and receive an answer, without insight into how that capability was built or could be extended. This matters because base models often already show some vulnerability-detection ability from training data, but that output is inconsistent, unstructured, and un-auditable, exactly the properties a hand-authored tool can guarantee.
This research introduces REMAP (Reverse Engineering MCP Authoring Platform), a containerized lab environment that teaches students to create their own Model Context Protocol (MCP) tools, the interface standard for how LLMs invoke external software, to support reverse engineering and vulnerability analysis. Each student deploys a single, fully containerized stack that requires no manual dependency installation: a chat interface, a sandboxed environment for developing and testing a custom MCP tool, and a headless instance of pyghidra-mcp for automated Ghidra-based binary analysis. A lightweight orchestration layer turns routine, error-prone operations, such as restarting a tool or selecting an active Ghidra project, into simple controls. A shared, centrally hosted model backend provides LLM inference for the whole class; when GPU infrastructure is unavailable, an optional local model container serves as a fallback, keeping the focus on tool authoring rather than model performance.
The accompanying curriculum guides students through three stages: a minimal "hello world" MCP tool that establishes the development loop, a standalone triage tool that flags decompiled code against common CWE patterns such as unbounded string copies, and a capstone exercise in which that same tool is chained with pyghidra-mcp's decompilation output to analyze real binaries. Rather than duplicating Ghidra's decompilation and cross-reference capabilities, students build a complementary tool demonstrating the value of custom tooling within an existing MCP ecosystem.
By containerizing the entire stack, REMAP eliminates operating-system and dependency variance common in applied AI labs, enabling reproducible deployment on student laptops or lab machines across platforms. This lowers the barrier to entry for cyber, computer science, and information systems students, who are not expected to be AI specialists, to engage with LLM tool-calling architecture while reinforcing reverse engineering and vulnerability analysis competencies. REMAP is being piloted in a graduate reverse engineering course this semester, with student performance on the capstone exercise and formative feedback on the workbench informing ongoing curriculum refinement. As Model Context Protocol and similar tool-calling standards become embedded in professional security tooling, REMAP addresses a growing gap between programs that teach reverse engineering and those that teach AI tool authoring, positioning graduates as authors of AI-assisted analysis rather than only its consumers.