Viral posts say AI agents can now rebuild any app or website. The evidence points to something narrower — and still useful.
AI agents can now explain and rebuild individual features, and sometimes simple apps. Nobody has shown them rebuilding a complex product — and for websites, the most important code is out of reach.
In early October an open-source project called REA went viral. A widely shared post, republished on KuCoin’s news site, said any commercial app feature could now be turned into “a perfect open-source implementation”5 and declared the “closed-source moat” debunked5. It reported nearly 80,000 GitHub stars in three days5. According to Runtime Wire, one commentator called the moment “Software is done.”6
Since then a stronger version has been circulating: that AI can now reverse engineer whole products and websites. We looked at what the tool produces, what has actually been demonstrated, and what a “whole product” really is.
REA connects an AI coding agent to established reverse-engineering tools. What comes back is assembly, pseudocode, references and maps of how modules connect3, with every conclusion split into what was observed, what was inferred and what is still unknown7. It isn’t the original source code. It’s pseudocode plus a list of things the agent couldn’t resolve8.
The project’s own advice is to give the agent a local target and one behaviour to investigate3. Even REA doesn’t claim to recover source code or clone apps automatically6.
The demonstrations are real, and they’re feature-sized. In a walkthrough published by Implicator.ai, REA examined a six-file Electron example and produced 21 findings in about 1.3 seconds, and a formatter rebuilt from those findings matched the original on all 14 test cases1. The project’s own showcase recovered one calculation from a 1996 game and checked it against all 63 compiled bytes9.
The most telling test went bigger. In a hands-on comparison reported by MindStudio, an agent rebuilt Mac’s Preview app twice — once with REA and once without. Both produced working clones, and the agent without REA finished first8. For a simple app, you can rebuild what you can see without reverse engineering anything. That’s impressive, but it isn’t the same as rebuilding a product.
For websites the limit is structural. Mozilla’s developer documentation notes that most of the code behind a dynamic website runs on the server4, and that’s where things like search results, recommendations and purchases are put together4. A tool working from the outside sees scripts, network requests and what happens in the browser3.
It can map the storefront. It can’t see the warehouse.
A convincing interface is the easy part. As Stork.AI’s analysis of AI attempts at rebuilding Photoshop put it, interface fidelity isn’t production capability, and decades of engineering are far harder to replicate than menus and panels10.
The agents doing the rebuilding are still unreliable over long jobs. On a benchmark of real computer tasks they now succeed about 66% of the time, up from about 12%, but they still fail roughly one attempt in three2. Across a large product, those misses compound.
Even a perfect copy of the code wouldn’t copy the business. The data, the customers, the integrations, the operations and the trust built over years don’t ship inside an app. That’s our view, and it’s the part the viral posts leave out.
REA’s own disclaimer says source-like output grants no rights to the original program or to the output itself11. As MindStudio notes, selling a close copy of a proprietary app raises real intellectual-property concerns whatever tools produced it8.
| The claim | Verdict | Why |
|---|---|---|
| AI can explain how a specific feature works | Yes | Demonstrated on real apps, with evidence for each finding1, 9 |
| AI can rebuild a feature or a simple app | Yes, often | Shown repeatedly — though a plain coding agent can often do it from the interface alone8 |
| AI recovers the original source code | No | The output is pseudocode, evidence and unknowns3, 8 |
| AI can clone a whole website | Front end only | Server code and data aren’t visible from outside3, 4 |
| AI can clone a complex product suite | Not shown | A convincing interface is not the same as the capability behind it10 |
| The rebuilt copy is yours to sell | No | No tool grants rights to someone else’s product8, 11 |
The real opportunity is narrower than the hype and still valuable. It’s now much cheaper to recombine existing ideas and produce a convincing demo8. Teams can inspect a desktop app before they adopt it, using an agent they already pay for1. And a focused subset of a big product, built around one user’s needs, has real appeal10.
The real risk is narrower too. Treat anything you ship inside a client app or web page as readable. Keep pricing logic, models, keys and business rules on the server. If your team uses these tools, decide up front who is authorised to analyse what — because that question now sits at the centre of this work6.
AI has made reverse engineering faster, cheaper and available to far more people. It hasn’t made products copyable. Features are now easier to understand and rebuild. Products — meaning the servers, data, operations and trust behind them — still have to be built.
Researched with AI assistance; every figure is linked to its source and reviewed by the team.
This article is commentary, not legal advice. Whether reverse engineering is lawful depends on your jurisdiction, the software’s licence and what you do with the results.
K3i has no affiliation with REA or the companies and publications named. Product names are trademarks of their respective owners.