Reverse engineering means working out how a finished product works without its design documents. Going into 2027, AI agents do much of the reading. The open-source REA project passed 460,000 GitHub stars by mid-November 2026, hosted AI models have started interrupting some sessions, and two legal milestones in 2027 are worth putting in your calendar.
| Fact | Detail | As of |
|---|---|---|
| REA GitHub stars | About 462,000, with about 113,000 forks | November 16, 2026 |
| REA repository created | April 14, 2026, first trending in early October 2026 | GitHub, Trendshift |
| Latest REA release we found | rea-agents 6.9.0, dated November 11, 2026 | November 16, 2026 |
| Licence and install | MIT licence, npx rea-agents setup | REA README |
| Native engines REA can drive | Hopper, Ghidra, IDA and Kuna | REA README |
| Agents REA setup supports | Claude Code, Codex, Cursor, Gemini CLI, Grok Build and others | REA README |
| LLM decompiler accuracy gap | 4.9% of outputs that passed every shipped test still behaved differently | arXiv, September 2026 |
| US DMCA section 1201 rulemaking | Opposition comments due February 22, 2027, new exemptions cover October 2027 to October 2030 | US Copyright Office |
| EU Cyber Resilience Act | Reporting duties live since September 11, 2026, full application December 11, 2027 | Regulation (EU) 2024/2847 |
In this post:
Reverse engineering is the analysis of a finished system to find out how it is built. The textbook version has three steps: pull information out of the system, build a model of it, and check the model against the real thing. In my reading, the last step is the one people skip most often, and it is where most wrong conclusions get caught.
| Area | What people are trying to do | Examples |
|---|---|---|
| Security | Analyse malware, find vulnerabilities, audit closed code | Ghidra, IDA, Frida |
| Interoperability | Make your software talk to a closed protocol or API | Samba, Wine |
| Legacy code | Recover rules nobody documented before a migration | Code reading, tests, interviews |
| Hardware | Recreate boards, chips or parts when the maker is gone | PCB and chip teardown, 3D CAD models |
| Devices you own | Repair, unbrick or change the behaviour of a gadget | Firmware extraction, smart TV and IoT work |
On the binary side, Ghidra is the best-known free tool. The NSA's Research Directorate built it, and Hopper and IDA are the other names you will hear most often.
A session runs as a four-step loop: survey the target, decompile it, name what you find, then test each theory by running the program. Every target is different, but this is the loop I keep seeing described for compiled programs, and it is the one an AI agent follows too.
Start with the file type, the strings inside it and the libraries it imports. A licence-server URL or an error message often points you at the interesting code within minutes.
The disassembler shows the machine instructions and the decompiler turns them into C-like pseudo-code. The result is close to the original logic and far from the original source, because variable names, comments and structure are gone and functions arrive with names like FUN_00401a30.
This is where the real work happens. Once you work out that FUN_00401a30 checks a licence key you rename it, every call to it becomes readable, and your model of the program slowly builds up in those names.
Run the program under a debugger or instrument it live with a tool such as Frida. If your theory says the function rejects keys shorter than 16 characters, feed it a 15-character key and see what happens.
REA, short for Reverse Engineer Anything, is an open-source MCP server and command-line tool that lets a coding agent such as Claude Code or Codex use reverse-engineering software directly, as InfoQ reported on October 17, 2026. MCP, the Model Context Protocol, is the standard way of giving an agent access to outside tools. The morluto/rea repository uses the MIT licence and was created on April 14, 2026.
The growth came late and fast. The repository sat mostly unnoticed until it reached GitHub Trending in early October 2026, passed 115,000 stars by October 12, showed about 409,000 when we checked on October 30, and showed about 462,000 stars and 113,000 forks when we checked again on November 16, 2026.
Through REA an agent can query a disassembler, inspect functions and cross-references, trace how a program behaves and test its own guesses against the binary. Analysis runs locally and each result comes with its evidence and its stated limits. REA's own pitch leaves out recovering the original source code: native analysis returns pseudo-code and assembly, and the agent writes new code from those findings.
The list of things REA can inspect has grown a lot since October. The README now lists more than 40 target types, and this is a selection:
| Target | How REA handles it |
|---|---|
| Native binaries | Uses an engine you already have: Hopper, Ghidra, IDA or Kuna |
| Electron and JavaScript apps | Static inspection of modules, imports, source maps and IPC with no native engine needed |
| .NET assemblies | Static inspection with no native engine needed |
| Android apps | Uses headless JADX and a full JDK, which you install yourself, plus built-in APK signature checks |
| Firmware images | Uses Binwalk or Unblob on Linux, which you install yourself |
| Websites and network captures | Browser observation, plus saved HAR and packet captures |
| Go and Rust binaries | Built-in reading of embedded module, crate and build information |
| Game and app packages | Unity IL2CPP metadata, Godot PCK archives, PyInstaller and Deno executables |
| Model files | Header and tensor metadata for GGUF and safetensors files |
Source: REA's README, checked November 16, 2026. Host support varies by target, so read the guide for yours before you rely on it.
The project's showcases give a feel for the range. They include tracing how Notion's Electron app handles the clipboard, rebuilding a sound-panning function from the game DX-Ball with tests that match all 3,205 original x86 cases, and recovering the bullet-ring angle maths from the PC-98 game Touhou 4. A newer use is building practice environments for AI agents: REA can turn a recorded web session into a local environment for reinforcement learning or agent testing.
Setup is one command, npx rea-agents setup, which shows a plan and backs up your existing configuration before it writes anything. It needs a recent Node.js (22, 24 or 26) and npm. It can install Hopper if you approve, but Ghidra, IDA, Kuna, Java and the firmware tools remain your job, and commercial engines need their own valid licences. REA changes quickly, so update before you report a bug.
Three things changed in the last weeks of 2026: REA widened its scope, hosted AI models began interrupting reverse-engineering sessions, and a fake copy of the project spread malware. All three affect how you should plan work for 2027.
On October 31, 2026, Tom's Hardware reported multiple user claims that Claude Code and the Claude Opus 5.5 model were blocking ordinary reverse-engineering tasks, including sessions that used REA. The same report notes that Anthropic introduced a Cyber Verification Program on October 6 that limits advanced cybersecurity capabilities in its frontier models to vetted organisations, and that Anthropic had made no official statement connecting the two (Tom's Hardware).
On November 10, 404 Media reported that the pattern covered OpenAI's Codex too. According to coverage of that report, one user said OpenAI banned their account for cyber abuse after they used Codex to reverse engineer Samsung TVs. OpenAI engineer Tarun Gogineni replied on X that the Codex blocks were mostly accidental, a result of boundaries that are hard to draw, and that Codex has no rule against reverse engineering (IBTimes UK).
REA's pseudonymous developer, Morluto, told Scientific American that if the big providers block him he will switch to an open-weight model, and community threads describe the same move towards local models (Scientific American).
What I take from this: the picture is still made of user reports and one informal reply, so treat it as unsettled. For planning purposes, assume a hosted model may stop halfway through a job. Keep your notes and renamed functions outside the chat, do not build a process that only works with one provider, and if your work is genuine security research, look at whether your provider offers a verification route. Trying to trick a safeguard puts the account at risk and is not something I would build a business process on.
On October 13, 2026, a user reported in issue #2424 on the official repository that a look-alike repository had copied the README and added a Windows standalone download. Write-ups of the issue say Microsoft Defender flagged the file as a trojan and that the fake has since been taken down. REA's README now states that the only official sources are morluto/rea, rea.tools and the rea-agents npm package. REA has no standalone Windows installer, so treat any REA .exe from elsewhere as malware.
Between our October 30 and November 16 checks, the README added a fourth native engine (Kuna), a long list of built-in static inspectors, and guides for building agent training environments from web and Linux desktop apps. It also carries a clearer responsible-use statement: REA is meant for lawful research, education, interoperability, software preservation and analysis of software you own or are authorised to examine.
I read through the top comments on the Hacker News thread, which had 814 points and 340 comments when I looked. Almost everyone agrees something changed, and the disagreement is about how big it is. I found four camps.
This is the loudest one. One commenter said there are now no excuses for rebuilding the most notorious closed-source binaries, from console system software to Nvidia's CUDA tooling, and the only obstacle left is the lawyers at those companies. Others asked the obvious follow-ups: what do software companies do if every program can be reverse engineered, do they all move to SaaS only, and how long before the backend of a bank or a hospital gets the same treatment. One long-time reverser said Richard Stallman would never have guessed that software freedom could arrive without anyone needing the source code.
Another commenter wondered whether REA is why their video feed filled up with vibe-coded clones of Photoshop, Illustrator, After Effects and Microsoft Office in the same week. That is a guess, and I did not find a source that ties those clones to REA. It does show how quickly the two stories got joined in people's heads. A widely shared guide on X by Miles Deutscher, dated October 21 and showing about 195,000 views on a mirrored copy, pitched REA mainly as a way to copy features from the apps you like.
One practitioner asked how REA is better than telling Claude to install a full reverse-engineering setup with Ghidra and get to work. Another said the top models already do this well on their own, and described feeding the Windows Remote Desktop client to Claude to fix two bugs that had annoyed them for years. A third praised the quality of REA's Touhou 4 decompilation but pointed out that the file layout seems tuned for AI use more than for mirroring what the original developers intended.
The most practical question was about refusals. Commenters noted that hosted models from the big labs tend to get nervous when you mention reverse engineering. That worry became the main story of the following month, as the section above describes.
The excitement is fair, because a free tool that lets any agent work through a binary was out of reach for most people a year ago. The "every program is now open source" version of the story is ahead of the evidence, in my view. REA's own showcases are a clipboard feature in Notion and decompilations of old games, its own pitch leaves out recovering the original source, and the study further down shows that agent output needs checking before anyone relies on it. Rebuilding a program from observation also leaves the original maker's licence and copyright where they were.
People are building three kinds of things: clones of commercial apps, ports and decompilations of old games, and fixes for devices they own.
The clearest public example of the first kind is ArtCraft. Vice, citing Ars Technica, reported that developer Brandon Thomas used Claude and Rust to build open-source clones of Adobe's apps, with names such as PhotoCraft, VectorCraft, FilmCraft and LightCraft, and described them as glitchy and still in early alpha (Vice). The project has grown since. The Adobe set covers seven apps, and in mid-October Thomas added WordCraft, GridCraft and DeckCraft, which imitate Microsoft Word, Excel and PowerPoint (Windows Central). The coverage I read found no public response from Adobe or Microsoft.
ArtCraft is a useful reminder that not every clone is reverse engineering. Neowin reports that Thomas describes the apps as clean-room implementations and says he did not decompile the originals or copy Adobe's source code. I found no source that ties ArtCraft to REA.
Clean-room engineering is the usual defence for projects like this. One side studies what the software does and writes a specification, and a separate side writes new code from that specification without ever seeing the original source. Courts have dealt with clean-room cases for decades, though never at the speed an agent can work.
The third kind gets less attention and may matter more to ordinary people. According to coverage of 404 Media's reporting, hobbyists have used agents to strip ads from smart TVs, turn smart picture frames into calendars and revive bricked gadgets. These are also the users who ran into the refusals described above.
Only after you have checked it, and this is where the "every program is now open source" idea runs into numbers. A September 2026 paper, When LLM Decompilers Recompile More and Preserve Less, evaluated eight LLM decompilation systems in nine configurations, and also tested 300 real GitHub library functions and 287 functions tied to known vulnerabilities (arXiv:2609.05370). Three results stood out to me:
The same habits apply to generated code in general. Our posts on the hidden risks of AI-generated code and common mistakes when building AI agents go deeper, and harness engineering covers how agents get wired up to tools like these.
I think this is the side of reverse engineering that matters most to ordinary businesses. Most will never open a malware sample, yet many own a system that works, that nobody fully understands and that has to move. Recovering what it does is reverse engineering with source code in place of a binary. It is also the safest place to use an agent, because you own the code and nobody's licence is in question.
Say a membership site has a custom module that applies discounts. The developer left years ago and nothing is documented, while the migration plan assumes standard pricing. Someone reads the code and finds special cases the business forgot it had: a discount for long-standing members, a cut-off date and an exception for one partner.
None of that shows up in a content export, so it has to be recovered before the move. Our Kentico 13 migration guide and common migration pitfalls cover the planning side.
It depends on where you are, why you are doing it and what you agreed to. US courts have treated reverse engineering for interoperability as fair use in cases such as Sega v. Accolade (1992), and the DMCA has a narrow interoperability exception. The EU allows decompilation for interoperability under strict conditions in Directive 2009/24/EC. A licence that bans reverse engineering can still bind you, as Bowers v. Baystate (2003) showed, and cloning a commercial product raises separate copyright and trade-secret questions. I am not a lawyer and this is general information, so check with one before you start.
The US DMCA exemption ruling. Section 1201 of the DMCA bans breaking access controls on copyrighted works, and every three years the Librarian of Congress, on the Copyright Office's recommendation, adopts temporary exemptions to that ban. The tenth round is running now. The Copyright Office published its notice of proposed rulemaking on October 28, 2026, with ten newly proposed classes of exemption. Supporting comments are due December 7, 2026, opposition comments on February 22, 2027, and replies on March 22, 2027. Whatever is adopted or renewed covers October 2027 to October 2030 (US Copyright Office). I have not gone through all ten proposals, so read the notice if your work touches access controls.
The EU Cyber Resilience Act. The Act's duty to report actively exploited vulnerabilities and severe incidents has applied since September 11, 2026, and the full set of requirements for products with digital elements applies from December 11, 2027. My reading of why this belongs in a post about reverse engineering: if you sell software or connected devices in the EU, more people can now inspect what you ship, and what they find can start a reporting clock for you.
These are my expectations for 2027, based on the last two months. They are opinions, and I have marked them that way.
Split the work by which side you are on: the company whose software can be studied, or the company that needs to study a system it owns.
Reverse engineering means taking something finished, such as a compiled program, and working out how it was built without the original plans. You watch how it behaves, split it into parts and write down what each part does. In software the usual goals are security analysis, compatibility with another system, or recovering logic that nobody documented.
REA is an open-source MCP server and command-line tool, released under the MIT licence, that lets AI coding agents such as Claude Code, Codex, Cursor and Gemini CLI use reverse-engineering tools like Ghidra, Hopper, IDA and Kuna. The agent can list functions, follow references, trace behaviour and test its own guesses against the binary. REA runs locally and does not claim to recover the original source code. It showed about 462,000 GitHub stars on November 16, 2026.
Install REA only from the official sources: the morluto/rea repository on GitHub, the rea.tools website or the rea-agents npm package. In October 2026 a look-alike repository offered a Windows download that antivirus tools flagged as a trojan. REA has no standalone Windows installer, so treat any REA .exe from another source as malware, and analyse untrusted files inside a sandbox.
Sometimes. Since late October 2026, users have reported Claude refusing or stopping reverse-engineering tasks, and one user reported an OpenAI account ban after using Codex on Samsung TVs. An OpenAI engineer said the Codex blocks were mostly accidental and that Codex has no rule against reverse engineering. Tom's Hardware reported no official statement from Anthropic as of October 31, 2026. Plan for a hosted model stopping mid-task.
It can do part of the job. Agents handle the survey, decompile and naming steps quickly, but a September 2026 study found that 4.9% of decompiled outputs which passed every shipped test still behaved differently from the original. A person has to confirm the agent's conclusions by running the program on real inputs.
It depends on the country, the purpose and the licence you accepted. US courts have allowed it for interoperability in cases such as Sega v. Accolade, and EU law permits decompilation for interoperability under strict conditions. A licence that bans it can still make it a breach of contract, so speak to a lawyer first.
Two legal milestones fall in 2027. In the US, the Copyright Office's tenth section 1201 rulemaking takes opposition comments until February 22, 2027, and the exemptions it produces run from October 2027 to October 2030. In the EU, the Cyber Resilience Act applies in full from December 11, 2027. Expect clearer AI provider rules on reverse engineering as well, though none had been confirmed by mid-November 2026.
Decompiling is one technique inside reverse engineering, and it turns compiled machine code back into readable, source-like code. Reverse engineering is the wider job of understanding what a system does and how its parts connect, using decompiling, debugging, network capture, documentation and testing.
No. REA helps an agent study a program and rebuild features from what it observes, but it does not claim to recover the original source code, and the results still need checking against the real program. Software rebuilt this way is new code, and the original maker's licence, copyright and trade-secret rights can still apply.
The people and documents that explained the system are often gone. Before a migration or rewrite, a team has to recover what the code actually does, including business rules nobody wrote down, and skipping that step is a common reason migrations miss features or break integrations.
If you are planning a migration for 2027 and the documentation stopped being true years ago, the first job is finding out what the system really does. Talk to our team and tell us what you are stuck with, and we will tell you honestly whether it is worth untangling or worth replacing.
Sources
REA project repository and README, morluto/rea (stars, forks and target list checked November 16, 2026): github.com/morluto/rea
Hacker News discussion, "REA: Reverse Engineer Anything," October 2026: news.ycombinator.com
Tom's Hardware, report on Claude blocking game cloning and reverse engineering, October 31, 2026: tomshardware.com
404 Media, "Come and Take It," November 10, 2026 (members only, read through secondary coverage): 404media.co
IBTimes UK, report on Codex and Claude restrictions, November 2026: ibtimes.co.uk
Scientific American, interview with the developer of REA, October 2026: scientificamerican.com
Vice, "Everything We Know About the Free Adobe Clones Built With AI": vice.com
Windows Central, report on ArtCraft's Microsoft 365 clones, October 13, 2026: windowscentral.com
Neowin, report on the ArtCraft Office clones and the clean-room claim, October 2026: neowin.net
Apidog, "What Is REA (Reverse Engineer Anything)?" (star growth to October 12, 2026): apidog.com
Liu, Raff and Micinski, "When LLM Decompilers Recompile More and Preserve Less," arXiv, September 2026: arxiv.org/abs/2609.05370
US Copyright Office, Tenth Triennial Section 1201 Proceeding, 2027 Cycle: copyright.gov/1201/2027
Travers Smith, note on the EU Cyber Resilience Act reporting requirements, September 2026: traverssmith.com