Letting an AI Agent Drive Ghidra: A Hands-On Test of REA’s 170 MCP Tools, One Recovered Password, and the Limits

The article explores the capabilities of REA, a platform that allows various AI agents to drive reverse engineering tools like Ghidra, Hopper, and IDA. The author tests REA's performance by analyzing a small C program they wrote themselves, which takes a password as input and checks it against a built-in constant. The program was built with both a symbolled and a stripped version.

The process begins with installing the rea-agents npm package, which took about 1 minute and 51 seconds. However, the installation process revealed that deep engines like Ghidra, Hopper, and IDA were missing initially. The author then proceeds with initializing the MCP (Multi-Component Protocol) server over stdio, which successfully returned 170 tools available for use.

The analysis phase involved several steps, including opening the binary, listing functions, pulling pseudocode, checking references, running the target, and diffing its behavior. The human's role shifted from merely paging through assembly code to asking the right questions and auditing the evidence chain. Each call to the AI agent returned a structured Evidence record containing the target file's SHA-256, provider information, raw and normalized results, confidence labels, and limitations of what was checked and inferred.

Despite the tool's potential, the author notes that the actual testing was limited to native ELF binaries, and no conclusions were drawn about other types of targets. The success of the password recovery, dependent on the correct key being 13 bytes, highlighted the limitations of the toolchain. While REA demonstrates promise in automating certain aspects of reverse engineering, the article emphasizes that human intervention and judgment are still necessary at critical steps.