AI NewsProductivityReported

One writer reverse-engineered seven devices in six months with an LLM

Herman used DeepSeek v4.1 Flash and Ghidra to reverse-engineer seven devices in six months. The write-up has 303 Hacker News points.

AI News

Editorial2 min read

LinkedInX

Why it mattersHardware reverse engineering used to take years. A single writer now does it in weeks by letting a model follow a strict hypothesis-testing loop.

Hardware reverse engineering used to be a hobby that took years for one device. Herman says he has done seven in six months.

Herman writes the Slow Chronicle blog under the handle ioma8. The post reached the front page of Hacker News with 303 points. Herman has been reverse engineering hardware and software for 20 years. He writes that the practice was manual until this year.

What he did

The post lists seven projects. Herman reverse-engineered the Alphasmart NEO typewriter. He mapped the serial protocols of the EmWave biofeedback device and Sony Alpha cameras. He wrote custom firmware for three music devices: the Arturia MiniFreak synthesizer, the Teenage Engineering OP-Z and the Akai MPC Sample sampler. He wrote custom firmware for the Xteink X4 e-reader. He fixed a broken macOS driver for an HP printer. He also cracked several abandonware games.

Each of those used to be a separate project that would take months of work on its own. Herman writes that he now finishes them one after another.

The loop he uses

The method is a strict hypothesis-testing loop. Herman tells the model what he thinks the binary is doing. The model proposes a change or an experiment. He writes the experiment, runs it, and reports the result back. The model updates its reading. He calls the move "hypothesis debugging" and says it stops the agent walking in circles.

The toolchain is named in the post. Herman runs DeepSeek v4.1 Flash as the model. He drives it from Ghidra through an MCP bridge, which lets the model read and comment on disassembly. He uses radare2 for the lower-level work. He writes his own firmware in Rust.

The practice points he lists are concrete. Each project gets its own folder of Markdown files, which the agent reads and writes. Every finding goes into the Markdown before the next step, so the agent can recover context after a crash. Herman says he refuses to let the agent guess: if a function is not understood, he adds breakpoints or logging and brings the real behaviour back into the chat.

Why a consultancy should read it

The lesson the piece carries is not that an LLM can read assembly. It is that one experienced hand plus a disciplined logging habit now covers work that used to need a team.

The limit is real. Herman names his tools because they work on his kind of binary. A commercial reverse-engineering job on a large modern application will hit limits he does not discuss. His bigger point still stands. A model that is not allowed to guess can take months out of a task that used to take years. The human has to record every step along the way.

Source

LLM assisted reverse engineering, by Herman (ioma8), Slow Chronicle. Hacker News discussion: 303 points.

Reported bySlow Chronicle

This item was written by an AI system from the linked source. Reveneau is responsible for what it publishes.

Share
LinkedInX
Start a project