号称能“逆向一切”的AI工具爆火，狂揽近30万颗Star，软件开发的护…
AI tool that claims to “reverse engineer anything” goes viral, racking up nearly 300,000 stars, a safeguard for software development…
Content Summary
Recently, the open-source project named REA (Reverse Engineer Anything) has quickly become popular on GitHub. It claims to enable AI to analyze software without source code.
Recently, the open‑source project called REA (Reverse Engineer Anything) has quickly gone viral on GitHub.It claims to enable AI to analyze a software’s internal implementation without source code and attempt to rewrite its functionality.
From the clipboard mechanism of the office software Notion, to the sound algorithm of the classic game DX-Ball, and then to the bullet‑hell calculation method of “Touhou Gensokyo,” a series of demos has attracted widespread attention.As of October 17, the GitHub page shows that its star count has exceeded 290,000.Tech blogger Matthew Berman even sighed on social media, “Software is done.”
Amid the hype, various AI reverse‑engineering software cases have begun to spread. Some have tried to rebuild Claude’s slide editor, others have used it to develop new control tools for their smart toilets at home, and developers have leveraged AI to launch open‑source applications that rival core software from Adobe, Microsoft, and others.
But as more developers personally experiment with REA, several issues have started to surface. In public testing on GitHub, some discovered that the AI would omit actual function calls yet still label the analysis as “complete”; sometimes it didn’t invoke REA at all, providing unverified answers directly. The REA authors also joined the discussion, responding to concerns about the tool’s invocation method and project design.
At the same time, another question has begun to arise: Those Adobe tools claimed to be “reverse‑engineered” by AI—were they really decompiled?
Reverse engineering has existed for a long time, why did REA suddenly explode in popularity?
To understand the role of REA, we first need to clarify how it differs from traditional reverse engineering tools.
Software reverse engineering is not a new technology. For a long time, engineers have been using professional tools such as Ghidra, IDA, Hopper to analyze programs. Even without source code, they can still study the program’s internal instructions, functions, and call relationships through disassembly, decompilation, and then infer how a particular feature is implemented.
However, this work usually requires strong professional experience. After software is compiled, variable names, comments, and some structural information from the original code disappear, and engineers must search for clues among a large number of instructions that are difficult to read directly, continuously verifying their judgments.
REA does not invent an entirely new decompilation technique.It mainly uses MCP (Model Context Protocol) to connect existing reverse‑engineering tools to AI programming agents such as Claude Code and Codex, allowing the AI to invoke tools, analyze results, and attempt to generate the corresponding code based on developers’ questions.
Image | REA reverse engineering workflow (source: REA GitHub)
Take the analysis of a software’s search function as an example. In the past, engineers had to open Ghidra themselves, locate functions, map call relationships, and then gradually organize the analysis results. Now, developers can ask the AI directly, and the AI, using REA, calls professional tools to find relevant code and clues.
This does not mean that professional tools can be replaced. The REA official documentation explicitly states that when using tools like Ghidra, developers still need to set up the appropriate runtime environment. REA itself also cannot guarantee the recovery of a software’s original source code.
But this way of having AI automatically invoke specialized tools happens to coincide with the rapid development phase of AI programming agents.
A REA recommendation post that appeared in the Reddit community received about 14.7 k upvotes. As related demos continue to spread, some people even posted specifically to ask what the sudden surge of AI reverse-engineering videos is all about.
However, developers quickly pointed out that REA’s actual capabilities may be overestimated: the real decompilation work is still done by professional tools like Ghidra, and REA mainly enables AI agents to use them. Moreover, similar interfaces have already appeared in tools such as IDA and Binary Ninja.
At the same time, the actual results gradually showcased by developers have made more people see the appeal of this workflow.
A developer tried to rebuild the Claude Slides slide editor. He first downloaded part of the editor’s program files, then used Codex with REA to analyze the obfuscated code. Because the code after the first round of processing was still quite messy, he used Claude Code to further organize and rewrite it, ultimately presenting a version that can perform some editing functions.
Another developer applied REA to his smart toilet at home. He used Codex and REA to analyze the Bluetooth communication protocol of the manufacturer’s WeChat mini‑program, and based on the analysis, he developed a computer menu‑bar remote control. In a public demo, he opened the toilet lid directly from his computer. According to him, the entire process—from protocol analysis to tool completion—took only about half an hour.
Image | A smart toilet control tool created by a developer after analyzing the Bluetooth communication protocol with REA (source: developer demo)
A similar effort is the open-claude-ui project on GitHub. The developer used REA to study the interface and interaction patterns of the Claude desktop app, cataloguing 30 UI components, 8 interaction features, and over 3,000 lines of design documentation, and released the related code.
These cases make the uses of REA more concrete: it can assist in researching web applications as well as help analyze the communication methods of smart devices. However, the publicly available results mainly focus on specific functions and interface modules, and do not yet constitute a full reconstruction of commercial software.
When some developers began testing REA with stricter methods, they also discovered issues that are not easily visible in the demo videos.
What did developers discover after testing it?
A developer posted a usage guide in the official REA GitHub Discussions. Unlike videos on social platforms that directly showcase reconstruction results, he chose a more verifiable approach: he selected a program with publicly available source code, let REA analyze it, and then compared the results with the original program.
He first chose the Linux compression tool bzip2. To increase the difficulty, he used a version that stripped debugging symbols, removing the information in the program that would normally make functions easy to identify. Then, he asked REA to locate the function responsible for outputting a certain prompt message and to analyze its call relationships.
The results were fairly smooth. REA found the target function, a function that calls it, and 24 functions that it calls. The developer cross‑checked with the original source code and found the results to be consistent.
But when switching to another program, the situation is different.
This time, he chose the esmtool program from the open‑source game engine OpenMW project, let REA analyze the call relationships of 50 functions, and then used the traditional tool objdump for verification.
The results show that the analysis missed actual call relationships for 22 functions. More notably, REA at the time still marked these results as “complete” and did not indicate any omissions.
During testing, he also encountered another somewhat unexpected issue:Sometimes the AI simply does not use REA at all.
He tried to have the local AI model invoke the REA analysis program, but in the first four tests the model never proactively used this tool, repeatedly trying other approaches and even reading the binary program as a regular text file.
After adjusting the workflow, the model began to provide answers, but still gave speculative responses. It wasn’t until he changed the process to run REA via a script first, then let the AI read the obtained analysis results, that all three tests produced the expected answers.
This set of tests also caught the attention of REA author Morluto. In a GitHub discussion he responded that REA originally used a prototype-style MCP tool‑calling design, but as AI clients gradually support on‑demand tool loading, this approach may no longer suit most agents.He believes that enabling AI to use specialized tools programmatically is still a direction worth exploring.
If those tests mainly focus on whether AI can correctly analyze code, other users have already started tackling more complex tasks, such as porting a classic game to a new platform.
A Reddit user wants to use Claude Opus to port the classic game “Thief 3: Deadly Shadows” to a MacBook Pro.
Image | Claude tries to reverse-engineer “Thief 3,” but only handles one function (source: Reddit)
At first, Claude told him that this work could take years. Then he gave Claude the REA repository. The AI downloaded Ghidra and decompiled a function in about a minute. But when it realized the entire game program contains roughly 32,000 functions, it did not continue to reconstruct the whole game.
Some people in the comment section think that AI is unwilling to continue performing complex tasks; others point out that even if you can quickly decompile a function, reorganizing tens of thousands of functions into a runnable game still requires a lot of work.
From these usage records, REA can already provide accurate analysis results for some tasks, but when dealing with large programs, complex call relationships, and AI tool scheduling, omissions or errors may still occur.In the broader AI software reconstruction boom, there are also some notable projects that don’t even use REA directly.
Has Adobe Really Been Reverse‑Engineered?
While REA is blowing up, software engineer Brandon Thomas has also drawn attention for an open‑source project.
Thomas has about 15 years of software engineering experience and previously worked at the payment company Square. Dissatisfied with the subscription pricing models of commercial software like Adobe, he used Claude to develop ArtCraft, an open‑source suite of seven applications that correspond to products such as Photoshop, Illustrator, and Premiere Pro. Shortly after, he launched WordCraft, GridCraft, and DeckCraft, expanding the target to Microsoft Word, Excel, and PowerPoint.
Image | ArtCraft offers multiple open‑source software alternatives to Adobe products (Source: ArtCraft)
These projects were quickly incorporated into the discussion of “AI reverse software.” However, in reality, Thomas did not use REA, nor did he publicly demonstrate the process of decompiling Adobe or Microsoft internal code. He calls his development approach “clean‑room reimplementation,” meaning he does not directly copy the original software code but builds applications with similar functionality from scratch.
So, how do these Adobe‑look‑alike programs actually perform? Joe Foley, an editor at design media outlet Creative Bloq, tested them personally.
To try out the photo editing tool LightCraft, he specifically photographed his cat, only to find that the software’s automatic selection could not accurately identify the subject. The Photoshop alternative PhotoCraft does support common formats like PSD, but some image‑processing features are still lacking; the video editing tool FilmCraft cannot handle certain 4K footage directly, and even after converting the format, it runs slowly.
Foley believes that these programs cannot yet replace Adobe’s mature products, but as early‑stage projects they already offer a certain level of usability.
Image | In actual testing, the automatic mask failed to accurately select the cat (source: Joe Foley)
In addition to functional shortcomings, ArtCraft’s development approach has also sparked controversy.
On Reddit, some people question why projects like ArtCraft describe themselves as “clean‑room re‑implementations.” Traditional clean‑room development usually emphasizes not directly copying protected code from the original software, and sometimes separates the analysts of the original program from the developers writing new code. Consequently, some commenters ask: when AI is involved in both software analysis and code generation, how can the independence of the development process be proven?
Some also argue that whether software is an open‑source project depends mainly on the openness of the code and the license, and should not be conflated with originality issues. At present, these discussions do not prove that ArtCraft has committed infringement, but the source of the code and the ownership of the results have become the focus of attention.
Similar issues have also affected the game reverse‑engineering community. In mid‑October, the decompilation project platform decomp.dev announced that it would stop publishing the project index and accepting new project submissions. Creator Luke Street said,AI has indeed increased decompilation efficiency, but some people have started using it to quickly generate code that hasn’t been thoroughly vetted, even borrowing years of work from other teams to prematurely release so‑called finished versions.The leaderboard originally meant to showcase research progress has also gradually caused some participants to focus too much on completion percentages while neglecting code quality.
In addition, the popularity of REA has been exploited to spread suspicious programs. A GitHub user reported a repository that impersonates REA, and the so‑called Windows installer it provides was flagged as a Trojan by Microsoft Defender, with eight detection engines on VirusTotal raising alerts. The associated archive showed 172 downloads at the time, but the actual number of infections could not be confirmed.
Behind these controversies is AI’s involvement in software development and reconstruction in various ways. Some people use REA to analyze existing programs, while others employ Claude to develop functionally similar alternatives from scratch. Although many projects have already demonstrated tangible results, as more users adopt them, issues such as incomplete functionality, uneven code quality, and intellectual‑property concerns are beginning to surface.
Going back to the REA that originally sparked the discussion, it does not invent a brand‑new decompilation technique, nor does it yet achieve AI automatically reconstructing arbitrary software. What it does is integrate existing professional reverse‑engineering tools into an AI agent, allowing the portion of analysis work that previously required engineers to do manually to be handled with AI assistance.
In just a few weeks, REA has amassed nearly 300,000 GitHub stars, bringing more attention to the previously niche field of reverse engineering. From developers rebuilding editors and controlling smart devices to the continuous real‑world tests appearing on GitHub, this toolset has demonstrated concrete use cases while also exposing its current limitations.
As for the popular claim on social platforms that “software is dead,” the existing cases are not enough to support such a conclusion. Today, AI can already use professional tools to engage in software reverse engineering and can also help developers quickly create functionally similar applications. However, figuring out how a piece of software works, rewriting the relevant code, and getting a new program to reliably perform the original tasks are still different matters.
Reference link:
https://github.com/morluto/rea
https://rea.tools/
https://www.reddit.com/r/BestGitHubRepos/comments/1x0s6lv/
https://www.reddit.com/r/OutOfTheLoop/comments/1x4h70o/
https://www.creativebloq.com/news/i-tried-those-free-ai-coded-adobe-creative-cloud-clones-and-they-have-some-big-problems
https://www.reddit.com/r/opensource/comments/1x78lqh/
https://listedarticles.com/articles/sunsetting-decomp-dev
https://www.gamesradar.com/games/the-community-aspect-of-decompilation-has-been-broken-by-a-wave-of-low-effort-ai-slop-and-now-a-major-community-tool-is-shutting-down/
https://github.com/morluto/rea/issues/2424
Layout: Hu Lihua
Note: Cover/featured image generated with AI assistance
Get MirrorSpeed