Maria Mendonca

Posted on Oct 17

What Happens When AI Makes Reverse Engineering Cheap

#discuss #ai #security #opensource

Two weeks ago, reverse engineering was a niche skill that security researchers put on their resumes. This week, it's the most popular thing on GitHub.

A project called REA (short for Reverse Engineer Anything) went from under 10,000 stars on October 6 to over 100,000 by October 12, with more than 30,000 of those on a single day. As I write this, its repo shows about 296k stars, 74.8k forks, 718 open issues, and 478 pending PRs. In the same week, one developer released free, open-source clones of Photoshop and Premiere, and then Word and Excel.

I have read this point multiple times in different tones and senses: "Software is done." Is it?

I spent 3 days reading through the coverage, the legal analysis, and the lobbying filings. The honest answer is that it's less dramatic than the hype, and more important than the skeptics think.

What REA actually is

Reverse engineering means working out how a finished program works without access to its source code. It used to take specialist tools (decompilers, debuggers) and days of patient work per feature.

REA is an open-source MCP server that connects AI coding agents like Claude Code or Codex to those same professional tools, such as Ghidra (released publicly by the NSA in 2019), IDA and Hopper, plus Chrome DevTools for web and Electron apps. You describe what you want in plain language ("figure out how this app's search works and show me the evidence"), and the agent investigates and reports back.

Two details are worth knowing before you repeat the hype:

It's not new technology, it's a lower barrier. Projects like GhidraMCP and IDA Pro MCP had been connecting AI to these tools for a year. REA's contribution is making it easy enough for ordinary developers to try.
Its own README is modest. It says it doesn't claim to recover original source code or clone applications automatically, and it doesn't ship a decompiler of its own.

Meanwhile, the UK's AI Safety Institute reported in April that GPT-5.5 solved a reverse engineering challenge in about ten minutes for roughly $1.73. Human professionals had needed about twelve hours for the same one. That's not about REA, but it shows where the cost curve is heading.

The headline case: ArtCraft

The project that put this on everyone's radar is ArtCraft, built by Brandon Thomas. On October 3 he announced free, open-source Rust rebuilds of Adobe's apps, then followed with Word, Excel and PowerPoint equivalents on October 7. His stated motive was frustration with Adobe's cancellation fees. (Adobe settled with the US government for $150 million in March over allegations about hidden early-termination fees, while denying wrongdoing.)

The details matter more than the headline:

It's not built with REA. Thomas says he used plain Claude Code, and that his repos explicitly forbid the AI from reading, disassembling or copying the originals' code. The agent may only study behavior, public file-format specs and appearance. That is the classic clean-room approach.
It's rough. The projects are labeled alpha. Reported problems include Photoshop files with Smart Objects misbehaving, crashes when importing lots of video clips, and only 5 of 30 filter tests matching Photoshop's output. A pixel artist called the Photoshop clone unusable. Thomas has already walked back his promise of full parity in a month.
Looking inside isn't always faster. In one test reported by MindStudio, a Mac app was rebuilt once with REA and once from screenshots alone. Both worked, and the screenshot version finished first.

So a clone that opens, edits and exports can now appear in days. A clone you'd trust with your livelihood still can't.

The good

1. Security research gets faster. The same capability that lets someone clone an app lets defenders analyze suspicious binaries, understand malware and audit closed-source software. Work that took a skilled analyst half a day now takes minutes. That's a real gain, though attackers get the same speedup (more on that below).

2. Pressure on pricing and lock-in. If a credible free alternative can appear within weeks, vendors can't rely on feature lock-in or cancellation friction. Buyers gain negotiating power, and the next license renewal looks different when an open-source option exists.

3. Preservation and interoperability. Reverse engineering has a long, respectable history: Phoenix's clean-room rebuild of the IBM PC BIOS in 1984 made PC clones possible, and Samba and Wine exist because of it. AI makes that work accessible to individual hobbyists. People are bringing orphaned and console-only games to new platforms, though that edge is where the legal trouble starts.

4. It rewards what's actually hard. If copying features becomes cheap, defensibility moves to data, integrations, reliability, support and trust. That's arguably healthier than competing on how long it takes to read machine code.

The bad

1. Piracy and takedowns. The same tools that port games also help people break copy protection, and studios are reacting. Activision pulled a video of an AI-reverse-engineered Modern Warfare 2, and Bethesda sent a cease-and-desist over an AI-generated browser game based on Fallout. Creators who thought they were doing a fan project found out otherwise.

2. Fewer places for vulnerabilities to hide. For decades, "nobody will bother reading our binary" quietly protected sloppy security. That protection is eroding. For defenders it's a chance to fix things; for attackers it's a faster way to find them.

3. The legal ground is unsettled. In the EU, reverse engineering a product you bought is allowed under trade-secrets law, and functionality itself isn't protected by copyright, though code is. Decompilation is only allowed under strict interoperability conditions. In the US, courts have found fair use in some disassembly cases (Sega v. Accolade) and in copying API declarations (Google v. Oracle). But there's a question nobody has answered: a classic clean room separates a human who reads the original from a human who writes the copy. A language model trained on enormous amounts of public code isn't obviously "clean." Earlier this year, a rewrite of the Python library chardet with Claude sparked a dispute over relicensing. As one lawyer put it, independent reimplementations may be defensible, but expect litigation once someone with deep pockets is affected.

4. Contracts are changing. Software vendors are updating their terms to ban AI-assisted reverse engineering and competitive benchmarking, and enforcing licenses harder. How enforceable those clauses are is still disputed, but the trend is clear: cheaper copying means more monitoring and more lawyers.

What about regulation?

This is where I want to be careful, because the story most people tell is tidier than the evidence.

It's tempting to say "companies are scrambling for AI regulation because of tools like REA." I couldn't find evidence of that link. What the filings do show is a record amount of money going into Washington:

AI-related federal lobbying reached a record of roughly $130 million in 2025, according to Bloomberg Law.
In the first half of 2026, OpenAI nearly doubled its federal lobbying to $2.22 million, and Anthropic nearly tripled its to $3.53 million, per the Financial Times. Meta alone spent $5.99 million in Q2, according to Axios.
Reporting describes the effort as aimed as much at deterring tougher rules, and shaping those around data centers and the testing of advanced models, as at pushing for them.

In other words, the lobbying is mostly about which rules get written (and avoiding a patchwork of fifty state regimes), not a panicked reaction to open-source cloning. There's also a December 2025 White House executive order pushing federal AI standards and aimed at limiting state regulation, and the pro-regulation side has its own well-funded groups.

Cloning is getting its own, quieter response: takedown notices, tightened terms of service, and lawyers. The thing worth watching is that none of this was designed with AI-driven reverse engineering in mind. The rules we have were written for a world where reading machine code was hard.

My take

As a frontend developer, I don't think software is over. What's over is the idea that having built the screens is a moat. The hard parts of a product (edge cases, accessibility, performance under real use, years of compatibility) are exactly what the clones still struggle with. If anything, this raises the bar for what's worth building.

What I'd actually take from this mini research:

If you sell software, ask what a competitor with an AI agent couldn't rebuild in three months. If the answer is "features," that's your next strategy conversation.
If you build with coding agents, write down what your agent may and may not look at. ArtCraft's repos do this in a plain text file, and it's a good habit regardless.
If you buy software, your leverage just went up. Alpha clones aren't replacements, but they change what a vendor can get away with.
If you use these tools, don't decompile software you don't have the right to analyze. "The AI did it" isn't a legal defense.

The tool changed. The legal and ethical questions didn't, and they're about to get a lot of testing.

Sources

I would like to know: What’s your take on AI reverse engineering? Is AI-assisted reverse engineering a gift to defenders, a gift to pirates, or both? Tell me in the comments.

I'm a Senior Frontend Engineer working on accessible, high-performance web apps. I write about what I learn as I build. You can find me on LinkedIn, GitHub and my Portfolio.

Top comments (3)

Carlos (Bloqarl) • Oct 17

I work on smart contract security, and we've kind of lived in this world for years. Everything deployed is public bytecode, so "unverified" never meant hidden. People decompile contracts all the time, attackers included. What changed for us is the speed, not the access.

My guess is it helps defenders more where the code was already out there, and attackers more where teams were counting on obscurity. Do you think desktop apps start moving the valuable logic server side because of this?

Maria Mendonca • Oct 17

This is a really good point. I personally feel the speed has increased for users and attackers alike, so it is moving both positively and negatively at the same time.

On desktop apps moving logic server side: I think some will, but it isn’t a full fix. REA’s README lists web apps and Android apps as targets too, and it can analyze captured network traffic, so the API is still observable. The protection shifts from “hard to read” to authentication with session tokens, rate limits on API calls, verification processes, abuse detection and terms of service updates.

Maria Mendonca • Oct 17

In simple terms, we can say AI-Assisted Reverse Engineering is a double edged sword, helping both attackers and defenders at the same time. It accelerates security research and analysis, but also enables the rapid creation of clones and potentially undermines software vendors' pricing and lock-in strategies.