· via Hacker News – Front Page (hnrss.org)
Veteran hacker says LLM agents compress years of reverse engineering work into months
A reverse engineer with 20 years of manual work describes an agent-driven pipeline — Ghidra skills, hypothesis logs, Rust probes — that turned multi-year projects into six months of output.

What happened
A blog post by a self-described veteran of two decades of hardware and software hacking, explaining how large language models now drive most of that work, surfaced on Hacker News's front-page feed on 18 November 2026. Writing on Slow Chronicle, the author says an agent-based pipeline has produced, in the past six months, results that each would have taken years of manual effort.
How the pipeline works
According to the post, the workflow starts by identifying the target's binary type, platform and build, then picking a suitable decompiler — although the author notes that current models cope reasonably well with raw x86 or ARM disassembly alone.
The agent is then told to record every finding in Markdown files as it works. The author credits this habit with three benefits: the notes aid the human's own understanding, they preserve context across model compaction or handoffs to a different agent, and they keep the agent from wandering in circles.
From there the agent maps the target's basic structure — main functions, key branching and imported functions — whether the target is a desktop executable, a website or a firmware image. That overview is usually enough for the author to start modifying the binary, building mods or simply understanding it better.
What to do when the agent stalls
The post lists several fallbacks. The agent can debug the target with breakpoints, or use logging, either enabled in the original binary or captured by an injected probe that writes to a file. The agent can be given outside access, such as computer-use controls for desktop applications or an API-driven emulator for firmware. The human can also order a different approach — following strings in the binary, for instance, when chasing cross-references has stalled — or read the full conversation and reasoning trace to spot when the model has latched onto a dead end, a judgement the author says still requires an experienced eye.
Skills that keep agents productive
Three agent skills anchor the workflow. One is Matt Pocock's "grill me", which makes the agent ask the human up to five questions, one at a time, when the answers would be decisive for progress. Another describes the reverse-engineering tooling itself — radare2 and a Ghidra MCP bridge — which the author says matters because agents otherwise abandon the task too early when lost; a linked repository includes 22 Ghidra scripts and a fix for GhidraMCP version mismatches. The third, written by the author and released publicly, is "hypothesis debugging": when stuck, the agent must write down a hypothesis about how the system works, design an experiment that definitively proves or disproves it, record the outcome in a CSV file and repeat. The author calls this loop, a variant of the scientific method, crucial to stopping endless guessing.
For implementation, the author prototypes probes in Python managed with uv, favouring fast, portable single-file scripts, and builds finished hacks in Rust, arguing that its strict compiler gives the agent immediate feedback on memory and syntax mistakes, at the cost of slower compile times.
Reported results
The claimed six-month tally is long: a hacked AlphaSmart NEO typewriter; fully mapped communication protocols for an EmWave biofeedback device and two Sony Alpha mirrorless cameras, enabling custom file-transfer and control apps and turning an A7 III into a wireless webcam; custom firmware for the Arturia MiniFreak, Teenage Engineering OP-Z and an Akai MPC sampler; firmware for an Xteink X4 e-reader; a macOS driver fix for an old HP printer; cracked licence checks on abandonware from the author's youth; and a lightweight custom Evernote client. The author also describes decoding a proprietary save-file format and building a personal Photoshop-style editor on top of it, arguing that the boundary between reverse engineering and ordinary software development has largely dissolved.
These figures are a single practitioner's self-reported estimates; the post offers no independent verification of the time saved.
Why it matters
The post is an anecdote, but a detailed one from someone with deep manual experience, and it illustrates where the value of expertise is moving: less grinding through assembly, more steering, auditing and redirecting agents. The process lessons — persistent notes for context, logged hypotheses, decisive experiments — apply well beyond reverse engineering to almost any agentic coding work. The same leverage also cuts both ways: interoperability, device preservation and abandoned-software maintenance get cheaper, but so does circumventing licence checks and probing the protections of locked-down hardware and software. For anyone building or defending such systems, the cost of someone else understanding them has just dropped sharply.
- #llms
- #reverse-engineering
- #ai-agents
- #security
- #ghidra