Industrial AI
AI can reverse-engineer your software. What protects your model?
Published · 4 min read

Reverse engineering used to take a specialist and a lot of patience. Tooling now hands much of that work to a coding agent. For anyone shipping a trained model to partner servers, customer sites or edge devices, that changes the economics of copying.
What REA is
REA, short for Reverse Engineer Anything, describes itself as reverse engineering with your coding agent. You install it, connect it to an agent and ask a question about an app, a game or a website. REA gives the agent tools to read the program, decode instructions, trace calls and observe what the software does while it runs. The agent then explains a feature or helps rebuild it.
The public example on the REA site walks through why Windows Calculator returns 220 for 200 + 10%. The agent reads selected x64 instructions from the installed library, follows the calls and recovers the rule. The project is open on GitHub.
What it is not
This is reverse engineering with a language model, not reverse engineering of a language model. REA is tooling for application, binary and runtime analysis. Nothing in its material claims to decompile model weights in general or to clone a trained model automatically, and it does not claim to break encryption.
The distinction matters. The point is not that one tool steals models. The point is that the skill barrier to inspecting a shipped program is falling, and a trained model shipped with that program is usually a file and a runtime away from anyone who looks.
Why this is a commercial risk
A vision model for defect detection or a classifier inside an OEM machine is often the one part of the product a competitor cannot rebuild from a teardown. Mechanics can be copied. Years of labelled data and tuning cannot, unless the result can be lifted from the box.
Today, most deployed models are decrypted into memory to run. Licence checks and dongles control who may run the model. They do not stop someone with access to the machine from reading the weights once they are in the clear. As inspection gets easier, contracts and good faith carry more of the load.
This is not hypothetical. A 2021 USENIX Security study by Sun and colleagues examined 1,468 Android apps with on-device machine learning and found that 41% did not protect their models at all. Of the apps that did use protection and whose ML functionality the researchers could trigger, they extracted the models from 18 of 29 using simple dynamic analysis.
Where encrypted model deployment fits
Wodan converts the model into an encrypted artefact and runs inference on it with fully homomorphic encryption. The weights never exist in plaintext on the machine, so there is nothing to copy off disk or dump from memory.
Encryption alone does not stop someone with unlimited plaintext outputs from training a copy of a narrow model by querying it. The owner-held key and query budgets are how we address that.
If your model ships inside someone else's machine, the useful question is simple: what does an engineer with an agent and a weekend find when they open the box? An architecture review maps that for one model and one target.
Related
Industrial AI