The state of AI reverse engineering
Figures as of 31 October 2026 · 490 projects · updates as the directory grows
People now use AI models to decompile games, read firmware, decode protocols and port old software. This report counts what the Reversed with AI directory records about that work: what gets reverse engineered, with which models and tools, and how often the result works.
Key findings
- 490 projects are recorded. 242 (49%) are reported as working, 168 as partial and 80 as research.
- Games and mods make up 60% of the directory, followed by firmware and hardware at 16%.
- Claude models appear in 403 of the 449 projects that name a model (90%), ahead of GPT / Codex with 56.
- Among 295 games, static recompilation is the most common method (134), with 47% reported as working. 36 aim for a matching decompilation, where the rebuilt code compiles to the original bytes.
- 33 projects (7%) connected the model to a tool through an MCP server. 65 used Ghidra and 4 used IDA.
- Only 42 authors (9%) say how long the work took. Of those, 14 finished in under a day, 14 in a few days, 8 in weeks and 6 took months or more.
What people reverse engineer with AI
Projects by the kind of target. The working share is the part the author reports as doing what it set out to do.
| Target | Projects | Works | Partial | Research | Working share |
|---|---|---|---|---|---|
| Games and mods | 295 | 125 | 136 | 34 | 42% |
| Firmware and hardware | 77 | 53 | 15 | 9 | 69% |
| Apps | 38 | 27 | 5 | 6 | 71% |
| Research and benchmarks | 27 | 7 | 1 | 19 | 26% |
| APIs and protocols | 24 | 16 | 5 | 3 | 67% |
| File formats | 11 | 9 | 1 | 1 | 82% |
| Legacy software | 8 | 3 | 4 | 1 | 38% |
| Malware (defensive) | 8 | 0 | 1 | 7 | 0% |
| Websites | 2 | 2 | 0 | 0 | 100% |
Which models people use, and how often it works
449 of 490 projects name the model they used. A project that used more than one family counts once in each. Version-level figures and benchmark results are on the model leaderboard.
| Model family | Projects | Works | Partial | Research | Working share |
|---|---|---|---|---|---|
| Claude | 403 | 211 | 145 | 47 | 52% |
| GPT / Codex | 56 | 20 | 16 | 20 | 36% |
| Gemini | 5 | 2 | 0 | 3 | 40% |
| Open or local | 13 | 4 | 1 | 8 | 31% |
Games by method and platform
The largest group. Method and original platform are read from each entry; the full list is on AI game decompilation.
| Method | Projects | Works | Partial | Research | Working share |
|---|---|---|---|---|---|
| Matching decompilation | 36 | 8 | 21 | 7 | 22% |
| Decompilation | 60 | 17 | 34 | 9 | 28% |
| Static recompilation | 134 | 63 | 59 | 12 | 47% |
| Port or rebuild | 34 | 17 | 13 | 4 | 50% |
| Mod or tool | 15 | 11 | 3 | 1 | 73% |
| Original platform (5 or more games) | Projects | Works | Partial | Research | Working share |
|---|---|---|---|---|---|
| PC | 63 | 33 | 22 | 8 | 52% |
| PlayStation 2 | 34 | 9 | 22 | 3 | 26% |
| PlayStation | 25 | 10 | 14 | 1 | 40% |
| Nintendo 64 | 23 | 15 | 6 | 2 | 65% |
| Xbox 360 | 21 | 10 | 9 | 2 | 48% |
| GameCube | 16 | 7 | 8 | 1 | 44% |
| Nintendo Switch | 15 | 4 | 9 | 2 | 27% |
| Xbox | 15 | 7 | 5 | 3 | 47% |
| DOS and home computers | 14 | 9 | 4 | 1 | 64% |
| Nintendo DS | 12 | 1 | 9 | 2 | 8% |
| Arcade | 8 | 7 | 1 | 0 | 88% |
| Game Boy Advance | 8 | 1 | 6 | 1 | 13% |
| SNES | 7 | 4 | 2 | 1 | 57% |
| PSP | 6 | 1 | 4 | 1 | 17% |
The tools around the model
The reverse engineering tools and agents projects name most often. General build tools and libraries are left out. Each tool has its own page in the tools directory.
| Tool | Projects | Working share |
|---|---|---|
| Claude Code | 136 | 58% |
| Ghidra | 64 | 39% |
| REA | 21 | 52% |
| Codex CLI | 15 | 53% |
| PS2Recomp | 9 | 22% |
| PSXRecomp | 8 | 63% |
| xboxrecomp | 7 | 43% |
| GhidraMCP | 5 | 60% |
| JADX | 4 | 75% |
Time and cost
Most authors don't say how long the work took or what it cost, so these figures cover only those who did. They are the authors' own numbers and are not measured by this site.
- Under a day
- 14
- Days
- 14
- Weeks
- 8
- Months or more
- 6
Time reported by 42 projects. 5 report a dollar cost: median $21, highest $91.
How these figures are made
- Every live entry in the directory is counted, 490 at the date above. Each links to its primary source.
- Outcome, model, tools, time and cost are what the entry records from the author's repo, write-up or post. The site does not rerun the work.
- The directory covers public, documented projects, so it leans toward work people chose to share. It is not a random sample.
- Game platform and method are matched from each entry's target, title and write-up, and a few games are hard to classify.
To cite: Reversed with AI, “The state of AI reverse engineering”, 31 October 2026, https://reversedwithai.com/state-of-ai-reverse-engineering. For a single project, cite its primary source.