How AI Models Now Reverse Engineer Apps Like Adobe and Premiere
AI models can now rebuild software like Photoshop without source code. Here's how reverse engineering benchmarks and clean room rules make this work.
What does it mean for AI to “reverse engineer” software?
It means an AI agent looks at what a piece of software does, not how its code is written, and then rebuilds that functionality from scratch. The agent never sees Adobe’s or Microsoft’s source code. It studies the behavior, the interface, the inputs and outputs, and writes new code that reproduces the same results. This has existed as a human discipline for decades under the name clean room engineering. What changed recently is speed and success rate. Coding-capable models are now solving reverse engineering tasks that used to take skilled human teams weeks, and they’re doing it in hours at low cost.
TL;DR
- Clean room engineering separates the people who study a product’s behavior from the people who write the new code, so no copyrighted source is ever copied, only the functional idea.
- A benchmark called SBench tests whether AI agents can recreate software functionality purely from observed behavior, with no access to the original source code.
- A model referred to as GPT-6 Astra reportedly solved a large majority of SBench tasks on the first attempt and nearly all of them within a handful of attempts, at low cost per task.
- People have already used this approach to rebuild creative software resembling Adobe’s Photoshop, with more attempts aimed at suites like Microsoft Office reportedly underway.
- The same technique applies to games, letting hobbyists port characters and mechanics across titles (Mario in Elden Ring, Minecraft mechanics in GTA) without touching the original studios’ code.
- In online multiplayer games, the same reverse engineering approach can expose or modify client-side assets, which raises serious fairness and anti-cheat concerns distinct from traditional cheat software.
- The legal footing depends entirely on whether the original source code was used; functional specifications are generally not copyrightable, but expect litigation to test the edges of this.
Built like a system. Not vibe-coded.
Remy manages the project — every layer architected, not stitched together at the last second.
How does AI reverse engineer an app without the source code?
The process mirrors human clean room engineering, just compressed. First, the AI agent observes the target software: how it responds to inputs, what its interface looks like, what outputs it produces for given actions. It never touches the original codebase. From those observations, it builds a functional specification, essentially a description of what the software does, not how it’s implemented internally.
Then, in a separate step, the agent (or a different instance of it) writes original code to satisfy that specification. Because the second stage works only from the behavioral description and never from copied code, the resulting software is, in principle, legally distinct from the original even if it looks and acts almost identically.
What’s changed is that a single capable coding model can now do both halves of this process at a scale and accuracy that used to require dedicated reverse engineering teams. The SBench benchmark, created by researchers from organizations including Google DeepMind, OpenAI, Anthropic, UCLA, UC Berkeley, and Columbia, exists specifically to measure this capability under “contamination free” conditions, meaning the model genuinely hasn’t seen the target source code during training or evaluation.
What is SBench and why does it matter?
SBench is a benchmark built around a simple test: give an AI agent only the observable behavior of a piece of software, no source code, and see whether it can reproduce that functionality faithfully. It’s designed to simulate real-world agentic cybersecurity and reverse engineering work, where an analyst or adversary typically never has access to proprietary code and has to infer behavior from the outside.
The benchmark’s research paper, titled something like “The Next Challenge for Agentic Cybersecurity: A Realistic, Contamination-Free Reverse Engineering Benchmark,” was first submitted in mid-August of this year. According to reporting on the benchmark, a model referred to as GPT-6 Astra solved a large share of tasks on a single attempt and nearly all tasks when given multiple attempts, at a notably low cost per solve.
The significance isn’t just the score. It’s how quickly the benchmark went from “hard, meaningful test” to “nearly solved.” That pattern has become common across AI benchmarks generally: a new eval is introduced, a frontier model saturates it within months, and the field moves to defining a harder one.
Is cloning software like Adobe or Microsoft Office actually legal?
The legal theory rests on the same logic that has applied to clean room engineering for decades: copyright protects the specific expression of code, not the general idea or function of a product. If a team (human or AI) only ever sees what a product does and never its underlying source, and then writes new code to replicate that behavior, that new code is generally considered an independent work.
This is why people experimenting with AI-driven reverse engineering are careful to describe the process as working from observed functionality rather than from any leaked or copied codebase. As long as that line holds, the legal exposure is much lower than straightforward piracy.
One coffee. One working app.
You bring the idea. Remy manages the project.
That said, this remains legally untested at the scale AI now makes possible. Courts have ruled on clean room cases before, but always in a world where the “clean room” side took weeks or months of expensive human labor, which naturally limited how often it happened. When a single model can run an entire clean room process in a couple of hours, the volume of clone products entering the market is likely to increase sharply, and that volume alone could invite new legal challenges from companies like Adobe or Microsoft.
What are people actually building with this right now?
Two categories have emerged clearly. The first is practical software substitutes. Someone recently used an AI agent, described as GPT-6 Astra running in an extended “ultra” mode, to rebuild a Photoshop-like application, reportedly as part of a broader attempt to recreate the full Adobe Creative Cloud suite. The interface reportedly looks close enough to the familiar Adobe layout that an experienced user could find every tool without a learning curve, though obviously rebranded. Reports suggest work has begun on reverse engineering Microsoft’s software suite using the same approach.
The appeal here isn’t complete feature parity. Most people use a small fraction of a mature app’s total feature set. An AI agent instructed to replicate “the layout and the 5% of functionality I actually need” can produce something usable in a matter of hours rather than months, without needing to match every edge case the original product handles.
The second category is entertainment and modding: taking characters, physics, and mechanics from one game and inserting them into another. Mario’s moveset dropped into Elden Ring. Minecraft’s building mechanics layered onto GTA’s open world. Spider-Man’s traversal system combined with Batman: Arkham Knight’s city. None of this requires the original studios’ code, just a model capable of inferring how the original mechanics work and re-implementing them in a new context.
What are the risks, especially in online games?
The most concerning application involves competitive online multiplayer games. Many of these games run a meaningful share of their logic, including 3D assets, rigging data, and rendering information, locally on the player’s machine, syncing only essential state with the server. AI-driven reverse engineering can be used to modify those local assets directly: making object textures partially transparent to see through obstacles, adding visual overlays that reveal other players through walls, or adding light sources and map information the game was never designed to show.
This differs meaningfully from traditional cheat software, which typically relies on memory injection or a separate program running alongside the game. Modifying the game’s local asset files directly means there’s no external script or second process for anti-cheat systems to detect, since the game simply loads the altered files as if they were original. That makes this approach harder to catch with conventional anti-cheat tooling and raises real questions about fairness and integrity in competitive gaming, well beyond the intellectual property questions that dominate the software-cloning conversation.
Frequently Asked Questions
What is clean room engineering?
It’s a method of recreating software where one team studies a product’s observable behavior and writes a functional specification, while a separate team builds new code from that specification alone, without ever seeing the original source code. This separation is what has traditionally kept the resulting product legally distinct from the original.
What is GPT-6 Astra?
Remy doesn't write the code. It manages the agents who do.
Remy runs the project. The specialists do the work. You work with the PM, not the implementers.
It’s referenced as a coding-capable AI model that achieved high success rates on the SBench reverse engineering benchmark, reportedly solving most tasks on a single attempt and nearly all within a few attempts, at a low cost per task.
Is it legal to clone an app like Photoshop using AI?
It can be, as long as the process never uses the original source code, only observed functionality and behavior. This follows the long-standing legal logic behind human clean room engineering, though the scale AI now enables hasn’t been fully tested in court.
Can AI reverse engineering be used to cheat in video games?
Yes, and this is one of the more concerning applications. AI can modify a game’s local asset files directly, rather than injecting memory or running a separate program, which can make the changes harder for existing anti-cheat systems to detect.
Does this mean software companies have no real competitive moat anymore?
If a company’s only advantage was the difficulty of building its software, that moat is weaker now. Companies that compete on data, network effects, brand trust, or ongoing innovation are less exposed than those relying purely on the complexity of their codebase.
