To thwart reverse engineering by agents, dev proposes blinding them

Now that LLMs can help analyze commercial software binaries and recreate their functionality, one software developer and student has created a project designed to make this process more difficult.

Timo Zhang's project is called Blind LLM Agents (BLA), and as its name suggests, it attempts to interfere with efforts by AI agents to scrutinize app code.

"Coding agents can reverse-engineer shipped packages, CLIs, and WASM in minutes," Zhang told The Register in an email. "Classic obfuscation was built for humans and static tools, not LLMs that reason about meaning. Blind LLM Agents is a protection layer for that threat model."

Such agents can also help developers build alternatives to existing commercial software.

Last month, Artcraft, a two-person development shop, released seven open-source apps that aim to be replacements for Adobe's commercial apps Photoshop, Illustrator, Premiere Pro, Lightroom, After Effects, InDesign, and Acrobat Pro. Artcraft claims these apps are "clean-room reimplementations" written from scratch in Rust using Anthropic Opus 5.5.

Setting aside the question of whether enlisting an AI model to reimplement an app really immunizes the creator from legal risk, those making apps face the possibility that someone will have an LLM probe published app binaries and reconstruct them through reverse engineering. There's in fact a framework for this, Reverse Engineer Anything (REA).

Zhang has faced some criticism on social media for creating a limited barrier to copying other people's work (which at this point might simply be the work of another LLM). But not everyone accepts that all publicly accessible binaries should be easily available for review and reuse. BLA cannot prevent committed reverse engineering efforts, but it may add a bit of friction to the process when LLMs are involved.

"It doesn't write app code — it hardens what you've already built," Zhang said. "A transform pipeline (plus MCP tools / an agent Skill) applies renaming, lane splitting, seed-derived string encryption, build polymorphism, WASM export renaming, decoys, self-seal integrity checks, and a coverage matrix that won't claim protection it can't prove. Behavior stays equivalent; recovery cost goes up."

Zhang said various evaluations like red teaming and CrackMe challenges to recover keys or flags from hardened artifacts show that BLA is effective.

"We also drive REA (a reverse-engineering agent) against those builds," he said. "Fail-to-break is the pass. CI tracks defense score, self-verifies integrity coverage, and checks functional equivalence. We raise the cost — we don't claim it is unbreakable."

The project website says as much, noting that it raises the effort for analysis and adds decoy code, but doesn't promise protection against traditional interactive reverse engineering tools like Ghidra or Frida.

Zhang said he's aware of other software projects that aim to interfere with AI, such as Poison Fountain for poisoning training data and Anubis for deterring bots and scraping. BLA, he said, is different. It's for hardening shipping code against reverse engineering and LLM-aided analysis. ®