Reverse engineering with AI means using artificial intelligence agents to analyze an already compiled program and understand how it works, without access to the source code. In October 2026 the topic exploded with REA, which connects agents such as Claude Code to tools such as Ghidra and IDA, and with the ArtCraft apps, presented as open alternatives to Adobe programs. The technique became faster and more accessible, but it still requires specialists, consumes a lot of tokens and doesn't copy what lives in a product's servers, data and partnerships.

Reverse engineering is the way back: starting from a finished product to find out how it was made. In software, it means analyzing an already compiled program, the file you install and run, to understand what it does inside without having the original source code: which functions exist, how data is stored, which servers it talks to and which rules it applies.
The practice is old and has legitimate uses. In the 1980s, Compaq and Phoenix recreated the IBM PC BIOS through "clean room" processes, in which one team documents the original's behavior and another writes new code without ever seeing the first. In 1992, Samba was born, letting Linux computers talk to Windows networks. Video game emulators, security tools and old-game preservation projects follow the same principle. In 2019, the NSA, the US security agency, released Ghidra for free, one of the main programs in the field.
In October 2026 the subject left its niche. Almost in the same week, REA went viral, an open tool that puts AI agents in charge of these programs, along with the ArtCraft apps, presented as free alternatives to Photoshop, Premiere and other Adobe products. Headlines talked about the end of Adobe and even the end of software. This guide brings together what is concretely known, based on the official repositories, academic research and the analysis developer Fabio Akita published on October 13, and explains what it changes for companies.
AI arrived in a discipline with decades of history, court cases and mature tools. The milestones that help explain the current moment:
When a program is compiled, code written by people becomes machine instructions. In that process, a lot is lost: variable names, comments, the organization into files. Taking software apart means trying to recover whatever is possible, in four steps:
The third step explains why famous projects take so long. Rebuilding the code of Super Mario 64 and Zelda: Ocarina of Time from the original games was done by volunteer communities over years, even though they are small, heavily studied programs.
It also explains why the current wave of old games ported to computers and browsers started before AI agents. According to Akita's timeline, it gained strength from 2019 on, with the free Ghidra and years of accumulated documentation. AI accelerated a movement that already existed; it didn't invent it.
None of the tools behind reverse engineering with AI is new. The main ones:
What changed is who can operate these tools and how fast. An AI agent chains dozens of calls, reads huge outputs, summarizes what matters and proposes the next step. It's work that used to take a specialist hours at every step and that can now be done, in part, by someone with less experience, or much faster by someone who already has it.
That has two effects. For defenders, auditing an application got cheaper. For attackers, too. And for the software market, the question that went viral was a different one: if AI takes programs apart and rewrites them, what stops someone from cloning any product? The next sections answer with what is known so far.
Language models are good precisely at the tedious tasks: suggesting function names, explaining what a piece of pseudocode does, writing tests and trying again when something doesn't compile. With AI agents, which use tools on their own, the model can operate Ghidra, read the result and decide the next step, in cycles that used to depend on a person at every step.
The gain is real, but it has an important limit. A study published on arXiv in September 2026, "When LLM Decompilers Recompile More and Preserve Less", by Chang Liu, Edward Raff and Kristopher Micinski, measured the effect of using AI to refine Ghidra's output on 300 real functions. The rate of code that compiles rose from 75% to 90%, but the rate of functions that behave the same as the original fell from 74% to 62%. In functions with known vulnerabilities, up to a tenth of the results erased the flaw without any warning.
In other words: compiling isn't the same as working the same way. AI produces plausible code faster, and plausible is exactly the hardest kind of error to notice. Tests and review by someone who knows the subject remain mandatory.
REA is an open-source project, published on GitHub under the MIT license, that connects AI agents such as Claude Code, Codex, Cursor and Gemini CLI to professional reverse engineering tools. It works as an MCP server, the standard that lets a model use external tools, which we explain in the guide to the Model Context Protocol. On October 19, 2026, the repository had passed 300,000 stars on GitHub.
In practice, the person installs REA with one command, points it at a file, app or website and asks the agent a question. REA runs the tools on the machine itself, such as Ghidra, IDA or Hopper for native programs and JADX for Android apps, and returns its conclusions along with the evidence and the limitations of each one. The project also analyzes JavaScript and Electron apps, .NET programs, firmware, network captures and SQLite databases.
What it doesn't do is just as important. REA doesn't decompile anything on its own: it coordinates tools that already existed. It doesn't recover the original source code and it doesn't copy apps with one click. Users need to have the file, install the tools (some paid, such as IDA and Hopper) and pay for the agent's tokens. According to Akita's analysis, one of the project's own examples, analyzing a SQLite database compiled to WebAssembly, took 42 tool calls and about 300,000 input tokens. The repository's warning is direct: AI-assisted analysis can come out incomplete or wrong.
We cover installation, costs and permitted uses in the article REA: what Reverse Engineer Anything is.
ArtCraft is a creative tools studio founded by American developer Brandon Thomas. In late September 2026, he started publishing free, open apps that follow the workflow of Adobe programs: PhotoCraft (Photoshop), VectorCraft (Illustrator), FilmCraft (Premiere Pro), LightCraft (Lightroom), EffectCraft (After Effects), DesignCraft (InDesign) and PdfCraft (Acrobat, launched as PrintCraft). According to TechSpot, Thomas said he wrote the apps in Rust with the help of Claude Opus 5.5, cited Adobe's cancellation fees as motivation and summed up the moment with the phrase "software is over".
The PhotoCraft repository describes the app as a clean-room reimplementation of Photoshop, built only from public specifications and observed behavior, with no Adobe code, shaders or files, and with no affiliation with the company. The license is MIT or Apache-2.0, at the user's choice. As of this writing, we found no independent verification of that claim.
The README itself is honest about the project's stage: PhotoCraft is in alpha and doesn't yet replace Photoshop in day-to-day professional work. The gaps the project lists include generative AI features, missing tools, typography depth and plug-in compatibility. Akita points out others: hundreds of open issues in the repository, limited color management for print (CMYK and ICC profiles) and incompatibilities with PSD files. The screenshot impresses; the engine is still being built.
We tell the full story, and the lessons for those who sell and buy software, in did AI clone Adobe? The ArtCraft case explained.

The central argument of Akita's analysis is that the value of modern software isn't all inside the file you download. And what isn't in the file can't be taken apart. He points to four barriers:

Cost matters too. Akita recalls the case of Bun, a JavaScript tool rewritten from Zig to Rust in 11 days with up to 64 Claude instances working in parallel, at an estimated cost of about US$165,000 in API pricing. It worked because there was complete source code, the original author in charge and a huge test suite to check every result. A clone made from a closed program has none of that.
The same reasoning explains why old games are being ported so quickly: they are small programs (Super Mario 64 is about 8 MB, versus about 150 GB for Baldur's Gate 3), studied for decades, with legal questions settled long ago and no further changes. Ports usually require the person to own their copy of the game.
More recent consoles also help: PS4, PS5, Xbox Series and Switch use x86 or ARM processors, the same as computers and phones, so the work focuses on translating system and graphics card calls. PS3 and Xbox 360, with their own architectures (Cell and PowerPC), are much harder.
Even so, there is hype. Akita notes that the community that ports games by hand already criticizes AI-generated ports for alpha-level quality, and that some headlines don't survive a test: a project announced as able to run much of the PS5 catalog ran, in practice, one indie 2D game.
A common argument is that, soon, tokens will cost almost nothing and cloning software will become trivial. Akita's analysis disputes it with numbers: the per-token price of frontier models fell between 1.2 and 3 times since 2023. What fell much faster, according to Epoch AI estimates he cites, was the cost of reaching the same quality, in the range of 5 to 10 times a year.
Even applying the most optimistic drop to Bun's cost, a clone without reference code or tests remains expensive. And there's a limit of another kind: according to GitHub's Octoverse report, 81.5% of contributions on the platform in 2025 happened in private repositories, code no public model saw during training.
Away from the headlines, the technique has legitimate and valuable uses, several of them listed by REA itself among the permitted ones:
In Brazil, no article of law expressly prohibits or authorizes reverse engineering. The Software Law (Law 9,609/98) protects programs as literary works, and the Copyright Law (Law 9,610/98) makes clear that ideas, methods and systems are not protected. The code is protected; the functionality isn't.
The risk shows up elsewhere: reproducing or selling third-party code (article 12 of the Software Law provides for detention and, when the reproduction is for commercial purposes, imprisonment of one to four years), bypassing copy protection, breaching the license agreement or using confidential information obtained through a contract or job, which the Industrial Property Law treats as unfair competition. In the United States, decisions such as Sega v. Accolade (1992) and Sony v. Connectix (2000) accepted disassembling programs for compatibility purposes.
We go deeper in is software reverse engineering a crime in Brazil?. This text is informational: for a concrete case, consult a lawyer.
Everything that runs on the user's device can be analyzed: the mobile app, the desktop program, the browser extension and your website's JavaScript. It has always been that way. The difference is that analysis now costs less time and less expertise, for attackers and defenders alike.
The most common weak point isn't the code itself, it's what gets put in it. GitGuardian's State of Secrets Sprawl 2026 report counted 28.65 million new keys, passwords and tokens exposed in public GitHub repositories in 2025, and commits made with Claude Code's help leaked secrets in 3.2% of cases, versus 1.5% on average. Not by chance, improper credential usage is the first item in the OWASP Mobile Top 10, the main reference for mobile app risks.
The full protection checklist is in reverse engineering and AI: how to protect your app.
If a product's edge is just a set of screens anyone can see working, it got cheaper to copy. Not free, as the stage of the ArtCraft apps and Bun's cost show, but cheaper. What stays hard to copy is what isn't in the code: proprietary data, services that run on the server, integration with the customer's process, service, brand and speed of evolution.
The flip side is an opportunity. The same technology that lets you analyze other people's software lets you understand your own: document old systems, find hidden business rules, test your own security and build internal tools faster. Deciding where to apply it is a strategic choice, and it's the kind of question we answer in our AI consulting.
There's no need to wait for the next viral case to act. Five steps cover both the risk and the opportunity:
The terms that come up most in the news about the subject:
Akita's conclusion is no. REA is a toolbox for specialists, not a copy button; the ArtCraft apps exist, but they are alphas; Adobe is still alive because much of its value lies in services and data; and, without someone who knows what they're doing, the tool delivers little more than a nice screenshot.
Our reading goes the same way. AI agents multiply the productivity of people who understand the subject and expose those who don't. For companies, the question isn't whether AI will replace the team, but whether the team knows how to use AI with judgment: review what it produces, protect data and choose where it pays off. That's why our AI training for teams teaches Claude, agents and automation with each team's real tasks.
It's using artificial intelligence agents to analyze an already compiled program and understand how it works, without the source code. The AI operates tools such as Ghidra and IDA, suggests names, explains snippets and writes tests, speeding up work that used to be almost entirely manual.
No. REA coordinates analysis tools and returns conclusions with evidence. It doesn't recover the original source code or clone apps, and the project itself warns that results may be incomplete or wrong.
Not yet. The PhotoCraft repository says the app is in alpha and doesn't yet replace Photoshop in day-to-day professional work. The other apps are listed as alpha or in development.
Analyzing a program isn't a crime in itself, but copying or selling third-party code, bypassing copy protection and breaching the license agreement can lead to civil and criminal liability. For a concrete case, consult a lawyer.
If you publish apps, programs or websites with logic in the browser, yes: review what's exposed, especially API keys and rules that should run on the server. And it's worth assessing where the same technique helps, such as documenting old systems.
In practice, yes. Installation is simple, but interpreting the results, validating what the AI concluded and knowing what you're allowed to analyze requires technical knowledge.
Talk to the SINOON team and find out which agent, squad or engagement makes sense for where you are now.