Vibeleaderboard
← All Intel
Intel / article

LLM assisted reverse engineering

Source
slowchronicle.com
Author
ioma8
Date

Context

This is one experienced hobbyist's account of how AI agents changed his reverse engineering, and its most useful part is a method for keeping an from going in circles. Reverse engineering means working out how a finished program or device works without its source code, usually by reading its compiled machine instructions. In his post, the author, ioma8, who says he has hacked hardware and software by hand for two decades, describes now mostly pointing an agent at a binary and letting it work. His main model is DeepSeek v4.1 Flash, which he chooses for speed and price, connected through an bridge (a connector that lets an agent operate another program) to Ghidra, a free disassembler and decompiler. Those tools turn machine code back into readable assembly instructions or approximate source code.

The method rests on making the agent write things down. He has it record findings in Markdown files as it goes, which he says keeps it on track after its is compacted or the work passes to another agent. When it gets stuck, his published hypothesis-debugging skill, a reusable instruction set for agents, forces a loop: write one hypothesis about how something works into a CSV table on disk, design an experiment such as a probe or added logging that will settle it, record the result, and repeat. He also gives agents a describing his Ghidra setup, saying that without one they often give up too quickly. And he reads the agent's full reasoning to spot when it has narrowed onto a dead end, then redirects it specifically, for example from following cross-references in the disassembly to following the text strings stored in the binary. For the final modifications he prefers Rust, partly because its strict compiler gives the agent immediate feedback on mistakes.

His results are his own and unmeasured. He lists projects from the past six months, including custom firmware for synthesizers such as the Teenage Engineering OP-Z, a mapped camera protocol that let him turn his Sony A7III into a wireless webcam, and a fixed macOS driver for an old HP printer, and says each would have taken him years by hand. He gives no timings or failure counts, and he credits his long experience for steering the agents. Measured evidence on agents reading binaries is mixed: ICALL-BENCH, a benchmark that tests whether models can work out which function a compiled program reaches through an indirect call, found program-level success rates from 23% to 81% across models and agents, with little gain from more information access on the hardest cases. The post is best read as a workflow from someone able to supervise it; it does not show how well agents reverse engineer without that supervision.

Terms in this piece · Glossary
  • AI agent — An AI system that doesn't just answer once but works toward a goal in a loop — taking actions, reading the results, and deciding what to do next.
  • MCP — The Model Context Protocol — an open standard that lets any AI assistant plug into any tool or data source without custom integration code.
  • context window — The maximum amount of text a model can consider at once — its working memory for the current conversation or task.
  • agent skill — A reusable instruction file that teaches an agent how to do one job well — the procedure, the tools, and what counts as done.
Read the source www.slowchronicle.com
Recommended reads
Comments

Checking sign-in…

Loading comments…