X Discussion: Open-Source BLA (Blind LLM Agents) Targets AI Reverse-Engineering Agent REA With Decoys and Resource Traps in Build Outputs, Claims 300x-2,800x More Tokens Burned, Nearly 3,800 GitHub Stars

Background & Industry Pain Points

Since October, the open-source AI reverse-engineering tool REA (Reverse Engineer Anything) has worried many developers: paired with a coding agent it can unpack Electron installers, recover SourceMaps and read WASM and local databases in seconds, mapping an app's core logic for well under a dollar. Classic obfuscation only renames variables, which LLMs read easily. Over the past day, many posts on X shared BLA (Blind LLM Agents, GitHub qybaihe/bla, MIT license), an open-source project aimed squarely at this kind of AI reverse engineering. An explainer post by the account codila drew about 5,700 likes and 650,000 views, and X's trend summaries also listed it. This is a community discussion, not a verdict from a security body.

Architecture Highlights & Mechanisms

According to the project README, BLA does not hide code; it makes AI reverse engineering misread and run out of budget. Poison: format-valid fake credentials, honey SourceMaps and realistic shadow state machines split the agent's context across conflicting leads. Exhaust: it deliberately trips REA analysis limits such as the 16 MiB ASAR header, the WASM interface budget and 32-level JSON nesting, downgrading analysis to limited or a capacity error. Entangle: JS, WASM and native sidecars verify each other with SHA-256 so decoys cannot simply be stripped. Decoy code sits behind always-false math conditions, and the project says all 14 test targets keep 100% input/output equivalence.

Authoritative Benchmarks & Measured Scores

There are two kinds of numbers. The first is the README table of live runs against real REA: after hardening Go buildinfo, Electron ASAR, Safetensors, GGUF and other formats, REA returns capacity or budget errors, and formats such as WASM are only partly decoded. The second is the widely shared 300x-2,800x token figure: it comes from a 14-target estimate table averaging about 640x token and 304x cost amplification, which the project itself labels internal estimates. No independent third-party retest has appeared, and manual workflows such as Ghidra or Frida are not covered, as the README states.

Developer Hands-on Guide

To try it, run npx bla-agents harden . in a release directory, or use npx bla-agents setup to register it with coding agents such as Claude Code, Cursor or Codex and let the agent run the hardening flow, or add the plugin to bundlers like Vite. Run a full regression suite before shipping and check that the size growth is acceptable, since some formats grow from tens of KB to several MB. The project says it is only for software you own or are authorized to protect, and commenters note that it raises the cost of reverse engineering but does not replace server-side checks and proper key management.