When AI Can Clone Your App: The Protection That Never Was

Thirteen apps on a Sunday

A few days ago I opened getartcraft.com/apps and spent a good while just scrolling. Thirteen apps: an image editor, a vector tool, video, photography, PDFs, motion graphics, page layout, interface design, word processing, audio, presentations, CAD and spreadsheets. Native for macOS, Windows and Linux, most of them also running in the browser. All open source. All free.

My first reaction was pure excitement. I've been a fan of free software for as long as I've been writing code, and here was something that looked like a complete creative suite that anyone, anywhere, could download without a credit card.

My second reaction took a few minutes longer to arrive, and it hasn't left since. This post is about that second reaction.

What ArtCraft is, and what it isn't

It's worth being precise here, because the details matter.

On October 7, Phoronix covered the project with a headline that sums it up: AI makes "open-source, clean room" implementations of Adobe Photoshop and Premiere. The ArtCraft team describes PhotoCraft, FilmCraft and LightCraft as clean-room reimplementations of Photoshop, Premiere and Lightroom, written from scratch in Rust. According to Phoronix's coverage, they were built mostly with Claude Code and AI agents.

This isn't a pirated copy. There's no leaked Adobe code in those repositories. The apps reproduce the experience: in the project's own words, "layouts, tools and shortcuts that working professionals already know, so there's nothing to relearn". Every app can also be driven by an AI agent through a command line, a JSON channel or an MCP server.

They're also honest about where they stand. PhotoCraft's README says it's in early alpha and "is not yet a Photoshop replacement for daily professional work". It's missing generative features, a handful of tools, depth in typography and plugin compatibility. And yet every Photoshop menu item is already wired to a command.

That last sentence is the one that kept me up. Not because it's finished, but because of how far it got, how fast, and how it got there.

The protection that never really was

For decades, a lot of us lived with a comfortable assumption: if you ship a compiled binary, an installer, an app in the store, your code is reasonably safe. Nobody's going to rebuild your product from a .exe.

That was never quite true. Reverse engineering is as old as software. In the early eighties, Compaq built an IBM-compatible BIOS using a clean room: one team studied how the original behaved and wrote a specification, and a separate team that had never seen the code wrote a new implementation from that document. Phoenix later sold clean-room BIOSes to anyone who wanted to build a PC. In 1992, a US court ruled in Sega v. Accolade that disassembling code to achieve compatibility could be fair use. In 2021, the US Supreme Court ruled in Google's favor for reimplementing the Java APIs. In Europe, the Software Directive explicitly allows decompiling a program, without the owner's permission, when it's indispensable to achieve interoperability.

So what actually protected companies wasn't the binary. It was the cost. Rebuilding something like Photoshop took years, specialized people and a lot of money. The obfuscation was a speed bump; the real wall was time.

AI is tearing that wall down.

Tools like REA (open source, MIT license) connect a coding agent to the same decompilers and debuggers security researchers have used for years: Ghidra, IDA, Hopper, JEB. You ask your agent how a feature of an app works, and it inspects the binary or the shipped code to answer. Their demo takes a version of Chrome's offline dinosaur game, recovers the exact acceleration rule from the code and rebuilds it in a small playable example. Their project includes an agent skill that covers native binaries, Electron apps, Android packages, firmware and web applications.

To be fair to REA: its documentation says it's for lawful research, education, interoperability, software preservation and analysis of software you own or are authorized to examine, and it explicitly rejects cracked tools and license bypassing. Most of the people who use this kind of tool are studying malware, keeping old software alive or making things work together. That's legitimate and valuable work.

But a tool doesn't choose who uses it. What used to take a team of specialists months now fits in a conversation with an agent. Research on LLM-based decompilation keeps improving year after year, and every new model is better at reading code it was never meant to read.

The two halves of my head

This is where I get stuck.

The half of me that believes in free software sees something wonderful. A student in Quito who can't afford a creative suite subscription gets professional tools. A small organization stops depending on a vendor. Knowledge that was locked inside a few companies becomes shared. For years I've argued that software should be something we build together, not something rented to us.

The other half of me looks at the numbers. According to its latest annual report, Adobe has 31,360 employees. That's 31,360 people with rent, kids, mortgages and careers built over decades. And that's just one company, the one that happens to be in this headline. Behind every product that an agent can now reimplement in a few weeks, there are teams of people who spent years designing, testing, supporting and improving it.

It's not that I want to defend a corporation. It's that I can't pretend the work of all those people is worth nothing just because a machine can now imitate the result.

It doesn't only happen to the big ones

What surprised me most is that the same machinery cuts in the other direction too.

In March of this year, the maintainer of chardet, a Python library downloaded millions of times, released version 7.0 as a complete rewrite made with Claude Code, from a design document, in an empty repository. With the rewrite, the license changed from LGPL to MIT. The original author of the project objected: in his view, it was an explicit violation of the LGPL, because the maintainer had known the original code for years, so it couldn't be a real clean room. No court has ruled on it.

Think about what that means. The same technique that allows rebuilding Photoshop without Adobe's code also allows "washing" the license of an open-source project. Copyleft, the tool free software has used for decades to protect the commons, depends on the code being recognized as a derivative work. If an agent can produce something functionally equivalent that doesn't count as a copy, that protection gets weaker too.

And there's another layer. In March, the US Supreme Court declined to hear the appeal over whether a work generated by AI can have copyright, which leaves in place the rule that it needs human authorship. So we're in a strange situation: code written mostly by an agent may not be protected by anyone, and code written by people can be imitated by an agent in a way that might not count as copying.

The question I keep asking myself

As a developer who has thought about building products, my question is very concrete: does it still make sense to have a good idea and launch an app?

Because if I spend two years building something good, and in a few weeks an agent can clone the experience, publish it for free and take away the reason anyone would pay for mine, what was the point?

The market has been asking itself the same thing. In February, software stocks went through what analysts called the "SaaSpocalypse": depending on the measure, around a trillion dollars in market value disappeared in a matter of days over the fear that AI agents would replace entire categories of software. Part of it recovered later, but the fear didn't go away.

I don't have an answer, but I do have some hunches about what still seems hard to copy:

• What never leaves your servers. If the valuable part of your product runs in your infrastructure, there's no binary to decompile. That's probably part of why so much software already lives behind an API.
• Data and network effects. You can clone an interface. It's much harder to clone ten years of data, or a community where everyone's already there.
• Trust. Support, security, stability, someone to call when something breaks. Companies don't only buy features, they buy someone who answers.
• Speed. If you can clone it in weeks, maybe the advantage is in being weeks ahead all the time, and in knowing your users better than anyone.

Even so, I'm not sure about any of this. All of those advantages are also being tested by the same tools.

What we don't know yet

The honest part is that there are more questions than answers:

• Is something truly a clean room when the model that wrote it was trained on half the internet, maybe including the code it's imitating?
• Can a law written for human teams in the eighties keep up with an agent that reimplements an app in a few weeks?
• If the license of an open-source project can be "washed" with a rewrite, what's left of copyleft?
• What happens to the people whose work was to build what can now be replicated?
• And on the other side: what happens to the people who never had access to these tools and now do?

Neither for nor against

I didn't write this to defend Adobe or to attack ArtCraft. I think what that project achieved is impressive, and I suspect it's just the first of many. I also think the companies and people who make software are facing something they had never faced before.

What I feel is closer to worry than to a stance. Worry about the present, because this is already happening and most of us are still assuming that shipping a binary protects us. And worry about the future, because I don't know what a good idea will be worth when copying it costs almost nothing.

Maybe the answer ends up being that what can't be copied was never the code. Maybe it's the people, the relationships and the judgment behind each decision. I'd like to believe that. But for now it's a hope, not a certainty.

If you've thought about this too, from either side, I'd really like to read you.

References and further reading

• Crafting Apps, the ArtCraft suite. getartcraft.com/apps
• Phoronix, "AI Makes 'Open-Source, Clean Room' Implementations Of Adobe Photoshop & Premier In Rust" (October 7, 2026). phoronix.com
• PhotoCraft, repository and project status. github.com/storytold/photocraft
• REA, reverse engineering for coding agents. rea.tools · github.com/morluto/rea
• LWN, "The relicensing of chardet". lwn.net
• Adobe, Form 10-K for fiscal year 2025 (headcount). sec.gov
• Forbes, "SaaSpocalypse Now? AI Is Disrupting SaaS, But Not All Software Is Doomed" (February 2026). forbes.com
• Directive 2009/24/EC, Article 6: decompilation for interoperability. legislation.gov.uk
• LLM4Decompile: decompiling binary code with large language models. arxiv.org