RSS
Tools

BLA (bla-agents): a CLI and MCP server that trips AI reverse-engineering limits on release files

BLA (qybaihe/bla) is a free MIT CLI and MCP server that adds decoys and padding to release files you own, so the AI reverse-engineering agent REA hits its limits. How it works, how to install it, and what it does not protect.

Padlock resting on a circuit board illustrating BLA protection against AI reverse engineering
Photo by Sasun Bughdaryan on Unsplash

BLA, short for Blind LLM Agents (GitHub: qybaihe/bla, npm package bla-agents), is an MIT-licensed CLI and MCP server that rewrites software you are about to ship so that AI reverse-engineering agents hit resource limits or chase decoys. It works on release artifacts such as JS bundles, Electron .asar archives, WebAssembly, SQLite, model files and native binaries, and is tuned against one open-source analyzer, morluto/rea. The disclaimer limits use to software you own or are explicitly authorised to modify. As of November 16 the repo, created October 25, had 5,786 stars.

Quick facts

  • What it is: a Node.js CLI (bla, bla-agents) plus a stdio MCP server the README says exposes 31 tools, and a bundled agent skill called blind-llm-agents.
  • Runs on: Node.js ^22.23.4 or 24 and later (package.json engines; the repo's .nvmrc pins 24). Runtime dependencies are only the Babel 8 parser, traverse, generator and types packages.
  • Licence and cost: MIT, free. bla.tools also has an Enterprise page; this article covers only the open-source package.
  • Latest version (as of November 16): 3.4.1, released November 15. There were 43 GitHub releases between October 25 and November 15.
  • Downloads: 18,051 on npm for November 8 to 14.
  • Formats: 26 by the README's count. JS/TS, Electron, WASM, SQLite, Safetensors, GGUF, Windows PE and others are marked verified against REA; Python, DEX, .NET, JVM and ZIP are designed-for only, PTX benchmark only.

What "Blue Defense" means here

The project casts itself as the blue team against REA-style agents that unpack an app and summarise its logic. The README groups the defence into four parts. Poison plants format-valid decoys such as fake enclave manifests and honey source maps. Exhaust pads files until the analyzer hits a hard limit, such as a 16 MiB ASAR header cap or a 32-level nesting cap in model metadata, so analysis stops as limited or with a capacity error. Entangle ties sibling files together with SHA-256 digests so stripping decoys breaks the app, but sidecarEntangle and self-seal are off by default, and for several formats the changelog says this is only checked host-side, not at load time. A runtime layer hardens Electron fuses and tries to detect debuggers or Frida.

The disclaimer says BLA never adds shell payloads or network callbacks. An optional --llm-directives flag adds text aimed at the analysing model; it is off by default and the skill file calls it risky bait. In the agent workflow, your coding agent also weaves decoy functions into real code paths and XORs string literals (which the docs say only raises grep cost), and BLA scores the result.

How to install

The README's quick start runs the published package with npx, and the installation guide recommends a dev dependency for CI release pipelines:

npx bla-agents harden .

# or, for a release pipeline
npm install --save-dev bla-agents

With 43 releases in three weeks and no npm beta tag, pin an exact version in CI, and install only from qybaihe/bla or the bla-agents npm package, as the README asks.

First setup and defaults to know

The installation docs describe only local npx and npm steps. Start with a dry run: the CLI docs say bla harden . --dry-run plans outputs while writing zero bytes, and bla lint . writes nothing and reports whether the strongest intensity would be refused. A real harden writes a .bla-hardened folder and a .blarc seed; the MCP tools refuse to overwrite files unless confirm_overwrite is set.

Three defaults matter. harden runs at bankrupt, the most aggressive of four intensities (stealth, balanced, maximum, bankrupt), with the biggest size growth. At maximum and bankrupt, every WebAssembly export not listed in --wasm-preserve-exports is renamed, breaking callers of a public WASM API. Discovery stops at 64 artifacts by default and marks the result truncated.

For agent use, npx bla-agents setup registers the MCP server and skill in Claude Code, Cursor, Codex, Gemini CLI and others by editing their config files; npx bla-agents uninstall removes them. Build pipelines can instead use the Vite, Rollup, Webpack or esbuild plugin, and native projects can generate C, Rust or Go headers with bla emit-headers.

What it does and doesn't protect

The effectiveness evidence comes from the project itself: a red-team report the maintainers generated on October 27 by running morluto/rea against their own hardened samples; there is no independent reproduction. In that report, for 10 of 13 formats REA's run ended in an error or budget rejection, two more hit or approached a cap, and GPU PTX still analysed normally. The README says this is not a claim about Ghidra, Frida, LLDB, Rizin or JADX sessions, which were not run, and that its token and dollar multipliers, around 1,069 times on its own 14-target benchmark, are internal estimates. The runs used Node 24; STATE.md notes SQLite is only capability-gated on Node 22.

Those trips come from size. In the report, a 607-byte ASAR grew to 16.9 MiB, a 54-byte WASM module to 6.56 MiB and a 761-byte JavaScript file to 116 KiB. The disclaimer also says no client-side scheme can beat a determined human analyst and that private API keys and signing keys belong on a server.

What stands out

  • It targets real limits: REA's limit constants are catalogued in source, and the changelog records which new REA commits were checked.
  • It publishes what went wrong. Issue #154 showed bla protect breaking direct-eval code while bla verify reported success; #155 showed one marker comment made an unprotected PTX file score as protected. Both were closed as fixed on November 9.
  • Public CrackMe challenges (v6 to v9) commit only SHA-256 digests of the keys, so anyone can point their own agent at them; STATE.md notes the published v9 build predates a later fix.
  • Each project gets its own seed in .blarc, and lint gates refuse bankrupt while plaintext FLAG-style secrets or single-point integrity checks remain.

How it was built with Claude

Of 545 commits on main, 16 carry a Co-Authored-By: Claude Code trailer, all from the maintainer between November 6 and 10, covering weave gates, a Babel 8 upgrade, SQLite and PE fixes and release cuts. Codex authored six October 25 parser fixes and Cursor Agent four website and README commits. The repo has an AGENTS.md and Claude plugin manifests but no CLAUDE.md. Most commits are by the maintainer, qybaihe, followed by DeryFerd.

Try it with your AI. This prompt has your coding agent try BLA on a copy of your release build at a moderate intensity, pausing before anything is written. Replace the three values first.

Help me test bla-agents on a copy of my own release build.

Build output folder: dist
Main file inside it: main.js
Test command: npm test

Rules: do not change my agent configuration, do not run bla setup, and do not use
--llm-directives, --in-place, --weave-decoys, --self-seal, --sidecar-entangle or any
--allow-* option. Do not delete anything. If a lint gate refuses, stop and show me the
finding.

1. Check node --version. bla-agents needs Node ^22.23.4 or 24+. If it does not match,
   stop and tell me.
2. Copy the build output folder to a new folder named bla-test and work only there.
3. Run npx [email protected] --help and npx [email protected] harden --help. Find the
   option that sets intensity; if there is none, stop.
4. Run npx [email protected] audit on the main file in bla-test; summarise the exposure.
5. Run npx [email protected] lint bla-test and show me the findings.
6. Run harden on bla-test with --dry-run and intensity balanced (not the default
   bankrupt). Show the planned outputs, where .bla-hardened and .blarc will be written,
   and whether discovery was truncated. Wait for my confirmation.
7. After I confirm, run the same harden command without --dry-run.
8. For each hardened file, run bla verify with the hardened file first and its untouched
   original from the build output folder second.
9. Point my test command at the hardened output, not at my source, and run it.
10. Report a table of file, original size, hardened size, verify result and test result.

Known limits

  • Cat and mouse: results hold against REA's current limits; a raised cap, another tool or a human analyst is not bound by them.
  • Verify with your own tests. Issue #154 showed hardening can change behaviour while BLA's verify passes; the disclaimer says to test protected builds in staging.
  • Some formats were weak until recently. In #176 a user found that hardening a plain C/C++ .so only appended about 1.16 MB of unloaded data; the maintainer merged a fix in PR #178 on November 12. Issue #157 found Windows PE hardening barely changed a model's token use; PE output was reworked in 3.2.3.
  • Signed APKs are refused, and setup edits every coding agent's config it finds unless you pass --client.
  • The red-team figures are the project's own, on a single REA version. Rerun the suite (npm run redteam:rea) or test your own artifacts before relying on them. Report problems on GitHub issues; SECURITY.md asks for a minimal reproduction.

FAQ

Which intensity does harden use? bankrupt, the strongest of stealth, balanced, maximum and bankrupt. Choose a lower one for a first test, because bankrupt gives the largest files, renames WASM exports and has the strictest lint gates.

Does BLA change how my app runs? The README says its 14 benchmark targets passed runtime-equivalence checks, but issue #154 showed behaviour changing while verify passed. Test protected builds in staging.

Do I need Wabt, tshark or Ghidra? No. The README says the shields are pure TypeScript; Wabt and tshark serve only the optional red-team suite.

How do I keep a public WebAssembly API working? List the exports with --wasm-preserve-exports (or wasmPreserveExports in .blarc); otherwise maximum and bankrupt rename every other export.

Source: GitHub: qybaihe/bla, BLA README, BLA disclaimer, BLA security policy, BLA installation guide, BLA CLI and MCP reference, BLA empirical REA validation report, BLA weave decoy notes, BLA blind-llm-agents skill, BLA AGENTS.md, BLA package.json, BLA CHANGELOG, BLA STATE.md, BLA CrackMe challenges, BLA releases, BLA commit history, Issue #154: JS protect changes runtime behaviour, Issue #155: PTX marker spoof, Issue #157: Windows PE harden results, Issue #176: hardened .so analysis, PR #178: mapped ELF64 .enclave for .so, npm: bla-agents, npm download counts for bla-agents, GitHub: morluto/rea

This story first appeared in the claudefox daily briefing for November 16, 2026.