A developer and student named Timo Zhang has built a tool that tries to stop AI helpers from taking apart finished software and rebuilding it. It is called Blind LLM Agents, or BLA. Zhang says it cannot stop a determined expert. The goal is to make copying slower and more costly.
What the problem is
When a company ships software, users get a compiled version of it, often called a binary. Reverse engineering means studying that version to work out how it behaves, then recreating it. AI coding agents can now help with that. Zhang says they can reverse-engineer shipped packages, command-line tools and web code "in minutes."
One recent example involves Artcraft, a two-person shop. Last month it released seven open-source apps meant to replace Adobe tools such as Photoshop, Illustrator and Acrobat Pro. Artcraft says they are "clean-room reimplementations" written from scratch with an Anthropic model. Whether that protects the creator legally is an open question, according to The Register.
What BLA does
BLA does not write software. It hardens software that already exists, and the app should behave the same afterward. Zhang says older scrambling methods were built for humans and basic tools, not for AI that reasons about meaning. The tool applies several changes:
- Renames parts of the code so names no longer hint at purpose
- Encrypts text strings inside the program
- Adds decoy code to mislead analysis
- Adds checks that detect tampering
What it cannot do
The project website says it raises the effort needed for analysis but does not promise protection against traditional hands-on tools such as Ghidra or Frida. Zhang put it this way: "We raise the cost, we don't claim it is unbreakable."
He says his own tests, including red teaming and puzzle-style challenges, show it works. He also runs a reverse-engineering AI agent called REA against protected builds, and counts a failure to break in as a pass. These are his claims. The material does not mention independent testing.
Why it is debated
Zhang has drawn criticism on social media for creating a barrier to copying other people's work. That work may itself have been produced by an AI model. Not everyone agrees, though, that every published program should be open to review and reuse.
What to know if you build software
For anyone who sells or ships software, AI-assisted copying is now a practical risk. Tools like BLA may buy time, but they are not a lock. Zhang also named other projects that fight AI in different ways: Poison Fountain targets training data, and Anubis deters bots and scraping. Pricing and how widely BLA is used were not stated.
The takeawayIf you ship software, AI copying is a real risk, and tools like this add delay but do not offer a guarantee.