You run product at a company that sells a professional desktop application. Your pricing rests on one assumption: a rival would need years and a large team to match what you ship. Last week tested that assumption three times.
On October 7 and 8, outlets reported that one developer had released free, open-source look-alikes of seven Adobe apps, which he says he built with Claude Opus 5.5. On October 9, a researcher published how a team of AI agents rebuilt the source code of a major shooter game from its compiled files. Meanwhile, a free tool that plugs coding agents into professional reverse engineering software passed 100,000 GitHub stars in about a week.
Before you read further, ask your CTO one question: which parts of our competitive logic run on the customer's device? Assume everything on that list can now be read.
Reverse engineering, working out how software functions without its source code, used to be limited by scarce expert time. AI agents paired with mature analysis tools now do much of that work, so the limit has moved to verification: can something check the agent's output? Where it can, complexity on the customer's machine stops being a moat. Advantage moves to data, server-side logic, workflows, trust and release speed.
Why is AI reverse engineering working now?
Software ships as a binary: the compiled file a computer runs, which people cannot easily read. Decompilation turns that binary back into readable code. The technique is decades old; expert time to make sense of the output was the scarce part.
Three things converged. Coding agents can now work on one task for weeks. The analysis tools are mature, such as Ghidra, the NSA's free reverse engineering suite. And some tasks now have a referee.
The referee is matching decompilation: rebuilding source code that compiles into a file byte-for-byte identical to the original. As TechSpot explains, that gives an agent "an unambiguous pass or fail." A wrong guess costs nothing; the check rejects it and the agent tries again.
Maurice Heumann, who ran the shooter project, put it this way: "A reliable verification harness that delivers an objective PASS or FAIL signal is the best feedback an agent can get." He also found that agents "have the desire to cheat if the assignment leaves room for interpretation." Some tried to edit the script that checked their work. With a strict check in place, cheaper models did the job.
So the bottleneck did not disappear. It moved from understanding code to checking it. TechSpot notes that graphics, physics, netcode and user interfaces have no referee, so code that looks right there can stay wrong.
What did the week actually show?
Commercial software: a rough Adobe suite, fast
Brandon Thomas, an Atlanta developer who says he has 15 years in the industry, released Artcraft: seven free, open-source apps written in Rust that mirror Photoshop, Illustrator, Premiere, Lightroom, After Effects, InDesign and Acrobat Pro. He says he built them with Claude Opus 5.5, citing frustration with Adobe's "cancellation fee," per HotHardware. Adobe agreed to a $150 million settlement with the US government in March 2026 over early termination fees, TechSpot reports. No Adobe response to Artcraft has been reported.
Thomas declared "Software is over" and promised full feature parity within a month, per TweakTown. He later said months, not years. Gizmodo's verdict: "janky for now, but it sort of works." Pixel artist Dan Fessler called it "completely unusable."
One number stands out. The project says Photocraft opens and resaves 134 of 135 real-world Photoshop (PSD) test files byte for byte, about 99%. Our read: file compatibility, long one of the stickiest forms of lock-in, is the part closest to done, because a file either round-trips or it does not.
Games: decompilation at scale, then takedowns
Heumann's project ran about three months, grew from four agents to more than 15 across Claude and Codex subscriptions, and used an estimated 600 to 700 billion tokens. He reports 99% of functions present and 83% byte-exact: roughly five in six functions provably identical to the original, in a game that plays. That took one researcher about one quarter. He does not name the game; Kotaku identified it as the PC version of Call of Duty: Modern Warfare 2.
Browser versions of Halo: CE, GTA Vice City and other classics are also circulating. Kotaku attributes them to AI decompilation; how they were built has not been independently verified.
Rights holders had already pushed back, Kotaku reported October 3: Activision had footage of the Modern Warfare 2 project disabled, and the creator of an AI-built browser game set in the Fallout universe dropped it after what he described as a cease-and-desist from Bethesda.
Tooling: reverse engineering as a plug-in
REA ("Reverse Engineer Anything") is an open-source MCP server under the MIT license. MCP, the Model Context Protocol, is a standard way to connect tools to AI agents. REA links Claude Code, Codex, Cursor, Gemini CLI and Grok Build to Ghidra, IDA Pro and other analysis tools, covering native binaries, Electron apps, .NET, Android, firmware and websites. Created in April 2026, it topped GitHub Trending on five of the seven days through October 12, per DEV Community.
For governance, the README says analysis runs locally, but "your agent receives the tool results, and its model provider has its own data policy." It also limits intended use to software "you own or are authorized to examine."
The limits
OpenAI reports that GPT-6 Astra fully solved 88% of SRE-Bench binary challenges on one attempt with a large inference budget, Popular AI notes. It also warns that "large, optimized, obfuscated, or server-dependent systems remain difficult."
Maintainers are drawing lines too: the open-source emulator projects mGBA and QEMU reject AI-generated code outright. Reading between the lines, byte-matching code reproduces the original faithfully, quirks included, which is not the same as code a team can maintain.
Where does this hurt, and where does it help?
Back to the product leader. The same capability cuts both ways.
Where it hurts
Moats and pricing. If logic runs on the customer's device, assume a determined rival can read it. Max Quimby's advice in AI Cloned Your App Overnight: "Treat client-side code as public documentation." We expect pricing that rests on feature count alone to feel pressure first, especially where customers already resent the terms. The moats that hold are the ones that compound beyond the code, which we cover in our AI wrapper product strategy guide.
IP and licensing. The law is unsettled and varies by jurisdiction. Popular AI's summary covers the key precedents: US courts in Sega v. Accolade and Sony v. Connectix found reverse engineering for compatibility to be fair use on those facts, while Bowers v. Baystate held that a license ban on reverse engineering could support a contract claim.
A clean-room rewrite, where one team describes what software does and a separate team writes new code without seeing the original, has been the traditional safe path. AI blurs it: when the chardet Python library was rewritten with Claude Code and relicensed from LGPL to MIT, its original author objected that a longtime maintainer cannot claim a clean slate.
Security. Attackers get the same capability. Obfuscation, deliberately scrambling code so it is harder to read, still raises the cost. Quimby again: "Obfuscation buys time. It no longer buys safety."
Where it helps
Legacy modernization. The same agents can read your own systems after the documentation and original engineers are gone. The referee rule applies here too: modernization goes well when you define what "correct" means, such as recorded production inputs and outputs, before an agent touches the code.
Buyer power. Our read: buyers gain bargaining power. With authorization, you can study what a vendor's black box does, estimate replacement cost and negotiate from facts.
How exposed is your advantage?
What we'd do this quarter
Inventory desktop, mobile, browser and firmware logic. Move what would hurt most if read to the server, where latency allows.
Source maps, debug symbols, embedded keys and internal endpoints make an agent's job easier.
Treat AI-assisted reverse engineering as a baseline attacker skill, and run an authorized analysis of your own builds first.
Use the exposure check above to find revenue that rests on feature parity. Ship faster so today's version is not the only target.
Decide who may run these tools, on what software, under whose authorization, and which model provider receives the results.
Are our anti-reverse-engineering license terms enforceable where our customers are? What would we need to show that a competitor copied us? How should our own teams document clean-room and AI-assisted rewrites?
Choose a poorly documented system, define a pass/fail check first, then pilot agent-assisted analysis.
Copying a product is not replacing a business
Artcraft shows how fast a product's surface can be imitated, and how far that sits from what professionals rely on: plug-ins, stability, support and a vendor who answers when something breaks. More of what you sell can be rebuilt now. The question is whether the rebuild differentiates anything.
What most teams will miss is the referee. Heumann's agents got good when he gave them a check they could not argue with, and the same holds for your own legacy code. At HatchWorks AI, our AI-Driven Development work on existing systems starts with the Brownfield Analysis Engine (scan, infer, validate, document), so the team knows what a system does before an agent changes it. For how deterministic tools and AI work together across a whole estate, see Modernization at Portfolio Scale.
