AI for reverse engineering your own software
To reverse engineer your own software with AI, run Kimi K3.1 Pro or GLM-5.3 in a coding agent next to a disassembler such as Ghidra.
Work that hosted assistants often decline
- Recovering a legacy binary whose source you lost
- Documenting your own device's protocol
- Understanding firmware on hardware you own
All of it is legitimate on systems you own or may work on, as why coding agents refuse explains.
Which model
| Model | Use it for | Price per 1M, in / out |
|---|---|---|
| Kimi K3.1 Pro | Large binaries; long outputs; screenshots of a disassembly view | $3.00 / $8.00 |
| GLM-5.3 | Text-only decompiled code; a second reading | $2.00 / $4.00 |
| DeepSeek V4.1 Flash | Bulk renaming across thousands of functions | $0.24 / $0.90 |
Silk Compute prices, read November 18, 2026.
Workflow
- Export the decompiled code
Run Ghidra on the binary and export its C output into the project.
- Let the agent name things
Ask for a name, argument types and a one-line summary per function, written to a file you review.
- Apply and iterate
Have the agent write a Ghidra script that applies the names, then repeat on the next layer.
- Confirm by running it
Check the agent's reading against the program's real behavior in a lab.
Licenses and local law decide what you may reverse engineer, so ask the owner first when the software isn't yours.
A prompt to start from
decompiled/ holds Ghidra's C export of our own legacy billing daemon, whose source was lost. Go function by function: propose a name, argument types and a one-line summary, and write them to names.csv. Then write a Ghidra Python script that applies names.csv.Paste it into Claude Code, Codex or OpenCode once it runs on an abliterated model.