Timo Zhang 1d3728ebeb Merge pull request #254 from qybaihe/claude/sad-jepsen-db4f40
feat: ship-ready agent journey, shield equivalence, honest reports — v3.5.0
2026-11-18 15:58:32 +08:00
2026-11-18 15:48:54 +08:00
2026-11-18 15:48:54 +08:00
2026-11-18 15:48:54 +08:00
2026-11-18 15:48:54 +08:00
2026-11-18 15:48:54 +08:00
2026-11-18 15:48:54 +08:00
2026-11-18 15:48:54 +08:00
2026-11-18 15:48:54 +08:00

English · 简体中文 · 繁體中文 · 日本語 · 한국어 · Türkçe · Русский · Tiếng Việt · ไทย · Deutsch · Español · Français · Українська · Polski · Português (Brasil) · العربية · فارسی · Bahasa Indonesia · Italiano · Ελληνικά · മലയാളം · Roman Urdu · Română


BLA Shield

BLA: Blind LLM Agents

Poison → Exhaust → Hallucinate. Make Reverse Engineering Bankrupt Again.

The open-source Counter-REA Blue Defense MCP & CLI that protects software releases against AI reverse-engineering agents (morluto/rea, autonomous decompilers, and LLM recon pipelines) — verified across 14 showcase benchmark targets.

npm version GitHub Trending MCP Server Formats Counter-REA MIT License

GitHub Trending — #5 Repository Of The Day

🌐 Official Website · 📖 Defense Guides · 📊 Live Benchmarks · 🎯 CrackMe Arena · 🏢 Enterprise · ⚖️ Disclaimer

Quick Start · How It Works · Empirical Results · Format Coverage · 14-Target Benchmark · CrackMe Arena · FAQ


Important

Defensive Use Only: BLA is engineered strictly for software authors and security teams protecting their own intellectual property and release artifacts. See DISCLAIMER.md. Always install from official sources: qybaihe/bla or npm i bla-agents.

Side-by-side: REA on an unprotected artifact vs BLA-protected output

🔴 Without BLA

AI reverse-engineering agents (REA + coding agents) unpack .asar archives, recover SourceMaps, trace IPC bridges, inspect .wasm / .sqlite / .safetensors, and clone core business logic in seconds for ~$0.20.

🛡️ With BLA (bla harden .)

REA's default pass stops on its own resource budgets (primitive-candidates > 256, ASAR > 16 MiB, WASM > Wabt cap, SQLite virtual trap), follows realistic shadow state machines, and projects 300×–2,800× modeled token amplification (token-estimator.ts). Most of those budgets can be raised by the caller — what that means.

⚡ Runtime-Equivalent by Design

Exhaustion graphs and decoy payloads sit behind O(1) number-theoretic opaque predicates (7x^2 - 1 \neq y^2 \pmod 7) or inert format sections. 14/14 bundled benchmark targets pass input/output equivalence tests (bla verify checks Counter-REA structural gates; run your project's test suite for custom artifacts).


⚡ 30-Second Quick Start

1. One-Command CLI Hardening

Point BLA at any project or release directory. It auto-discovers build outputs (dist/, build/, .asar, .wasm, .sqlite, .safetensors, native binaries, Python/C/Rust/Go) and runs audit → protect → verify:

# Node.js ^22.23.4 or >=24.0.0 recommended
npx bla-agents harden .            # writes a partial mirror to .bla-hardened/ — nothing ships yet
npx bla-agents harden . --apply    # ship it: replace the build output with the hardened files
npx bla-agents harden . --restore  # undo: put the original build output back

Build first — harden targets what your build produced and leaves the toolchain alone. --apply only overwrites build output (never source) and keeps the originals outside the project; run your own tests against the promoted build before releasing.

2. Agent-Driven Closed-Loop Defense (MCP + Skill)

Register BLA once with your coding agent (Claude Code, Cursor, Codex, Windsurf, Gemini CLI, GitHub Copilot CLI, VS Code, OpenCode, Antigravity):

npx bla-agents setup

Then simply prompt your coding agent:

Harden my software against REA

Your host agent automatically orchestrates BLA's MCP tools (bla_weave_plan → source-level logic weaving & string encryption → bla_verify_weave → bla_harden → bla_verify) to harden both your source architecture and compiled artifacts. Every result carries nextSteps, so the agent finishes the job: it asks before promoting the hardened files over the build output (apply), then runs your tests.

3. Build Bundler Plugin (Vite / Rollup / Webpack / esbuild)

Protect JS chunks, strip/replace SourceMaps with VLQ v3 honey maps, and shield .wasm assets automatically at bundle time:

import { defineConfig } from "vite";
import { blaVitePlugin } from "bla-agents/bundler";

export default defineConfig({
  plugins: [blaVitePlugin({ intensity: "bankrupt", enableHoneySourceMaps: true })],
});

🧠 How BLA Works: Agent-Driven Closed Loop

Traditional obfuscators only rename variables (_0x1a2b), which modern LLMs effortlessly summarize, or wrap binaries in heavy VM packers that break cross-platform builds and degrade performance.

BLA 3.x combines format-native deterministic shields with host-agent surgical source weaving:

flowchart LR
  subgraph S1["1 · Recon & Plan"]
    A["bla_audit / bla_lint"] --> B["bla_weave_plan"]
  end
  subgraph S2["2 · Agent Source Surgery"]
    B --> C["bla_encrypt_strings\n(AST Span XOR + Keystream)"]
    C --> D["bla_bind_scope\n(Multi-Site State Entanglement)"]
    D --> E["bla_verify_weave"]
  end
  subgraph S3["3 · Artifact Shielding"]
    E --> F["bla_harden / bla_defend\n(19+ Format Shields + Self-Seal)"]
  end
  subgraph S4["4 · Counter-REA Gate"]
    F --> G["bla_verify\n(Structural Counter-REA Checks Only)"]
  end
Defense Pillar Mechanism Effect on AI Reverse-Engineering Agents
1. Poison (Deception) Plants format-valid canaries, SGX/TPM2/Secure Enclave delegation manifests, VLQ v3 honey SourceMaps, and plausible shadow state machines. Forces LLM agents to split context across conflicting hypotheses and chase synthetic rabbit holes.
2. Exhaust (Budget Breakers) Trips the analyzer's default limits (primitive-candidates, alias-selections, ASAR 16 MiB header, WASM Wabt representation, SQLite virtual column eval, Model 32-level nesting, PCAP tshark field pass (fixed 8 MiB), Electron 32 fuse sentinels). Downgrades REA's default examination from COMPLETE to LIMITED / CapacityError. Where REA lets the caller raise a limit this costs the attacker a retry, not the analysis — see the last column of the results table.
3. Entangle (Anti-Strip & Self-Seal) Binds sibling artifacts (JS ↔ WASM ↔ Native sidecars) via runtime SHA-256 digests (--self-seal, --sidecar-entangle) and multi-site caller scope folding (bla_bind_scope). Prevents red-team agents from stripping decoys, nop-patching single-point checks, or calling exported functions as offline oracles.
4. Runtime Truman Show Hardens Electron V1 Fuses (RunAsNode=0, Inspect=0) and detects CDP / Node Inspector / Frida / WebDriver attachment (runtime-shield.ts). Serves believable synthetic telemetry during live debugging while keeping normal user execution untouched.

🔬 Empirical Results vs. Real REA

These numbers come from running real morluto/rea against BLA-hardened samples on commodity hardware (npm run redteam:rea). Full machine-readable report: docs/redteam/empirical-rea-validation.md / .json.

How to read this: “Empirically verified” means the live REA analyzer path returned the listed budget / capacity / redaction failure in this environment. It is not a claim about Ghidra/Frida/LLDB/Rizin/JADX sessions we did not run here.

Most of REA's budgets are defaults the caller can raise. As of REA 05be423 (2026-11-17) its limits are per-call inputs or settings, and REA's own error tells the agent which one to raise (asar.max_header_bytes up to 256 MiB, max_interface_bytes up to 512 MiB, max_chunks, semantic_precision). A ❌ below therefore means REA stops here as it runs by default: a size-based plant costs an attacker one retry, it does not lock REA out. The last column says which limits REA keeps fixed, and JavaScript strings is the one row measured with the budgets raised.

Empirically Verified on Commodity Hardware

Last recorded 2026-11-18 against REA 05be423 under Node v26.12.1 (macOS arm64) with Wabt 1.0.42 and tshark 4.4.1. REA analyses packet captures only on Linux / Windows, so on macOS the harness runs REA's two tshark passes, with the arguments read from REA's own provider, and hands their output to REA's parser. Soft-skips keep Node 22 / CI green when external native tools are absent; a probe that cannot call REA at all is reported as a harness error (status: unverified, exit 2), never as a budget hit.

Format Before → After REA Outcome After BLA Exact REA Budget / Error Triggered Can the caller raise it?
Go buildinfo 50 KiB → 1.17 MiB ⚠️ Rejected as malformed BinaryFormatFailure — Go build-info string range is truncated. REA reads no build info from the hardened file, but it now checks the string range before max_build_info_bytes (1 MiB), the budget this plant was sized for Not measured
Erlang BEAM 84 B → 241 KiB ❌ Capacity error max_chunks — BEAM form exceeds the 64-chunk budget Yes — max_chunks
EVM bytecode 16 B → 4.00 MiB ⚠️ Admitted, just under the cap 4,190,208 of 4,194,304 bytes of EVM_PRODUCER_LIMIT_DEFAULTS.input_bytes; the EVMole worker itself is not run by the harness Yes — producer_limits.input_bytes
PCAP 24 B → 1.91 MiB ❌ Budget rejected REA asks tshark for 25 selected fields and holds that output to the 8 MiB TSHARK_JSON_MAX_BYTES: the hardened capture prints 12.7 MiB, so inspect_pcap_capture fails with a non-retryable resource error. Ten DNS responses do it — 4,076 answers each, every one naming a 251-character owner through a 2-byte compression pointer, which tshark prints in full. Without them REA read the 160 other decoys (116 KiB) No — fixed; REA has no setting for it
Electron ASAR 607 B → 16.9 MiB ❌ Capacity error DEFAULT_ASAR_MAX_HEADER_BYTES (16 MiB) Yes — asar.max_header_bytes, up to 256 MiB
WASM + Wabt 54 B → 6.56 MiB · 251 exports ⚠️ ifaceState=partial max_interface_bytes (32 MiB) — Typed WASM interface representation exceeds max_interface_bytes Yes — up to 512 MiB
SQLite (Node 24, --sqlite-eval-trap) 12 KiB → 56 KiB ❌ Eval trap on the honey tables SqliteColumnEvaluationFailure — unknown enclave_unseal_payload() on virtual trap columns. Opt-in (the default through 3.4.3): it also stops the application's own PRAGMA integrity_check / VACUUM. Default hardening adds honey tables REA can read; application tables are never hidden. Not a budget — an evaluation failure
JavaScript 761 B → 153 KiB ❌ Budget rejected alias-selections + getter-expansion + primitive-bytes + primitive-candidates Yes — semantic_precision
JavaScript strings 340 B → 154 KiB ✅ 0 of 3 known values in REA's result An endpoint URL, an IPC channel and a SHA-256 digest, at 1× / 4× / 64× budgets; REA returns all three from the unhardened file. Through 3.4.3 the pieces were written in clear: REA listed the channel at its defaults and all three once the budgets were raised 4× Raised here, to 64×
Safetensors 128 B → 9.7 KiB ❌ Capacity error MODEL_JSON_NESTING — Safetensors JSON nesting exceeds 32 levels (HF-loadable; tensors intact) No — fixed at 32
GGUF 24 B → 7.3 KiB ❌ Capacity error MODEL_METADATA_NESTING — GGUF metadata nesting exceeds 32 array frames No — fixed at 32
Electron fuses 50 KiB → 1.17 MiB ❌ Budget rejected ELECTRON_FUSE_LIMITS.maximumSentinelsPerSlice (>32) No — fixed at 32
Flutter pool 50 KiB → 1.17 MiB ⚠️ Projected cap MAX_POOL_URIS / MAX_POOL_TOKENS (REA truncates with “at least N”) Yes — uri_limit / token_limit

Reproduce locally:

export REA_WABT_BIN_DIRECTORY=$HOME/opt/wabt-1.0.42/bin   # optional
npm run build && npm run redteam:rea

npm run verify:rea checks BLA's catalogue of REA limits against the same checkout (74 checks at 05be423) and says which of them are per-call defaults.

Designed-for paths not empirically run on this machine (click to expand)

These shields exist in-tree and are covered by unit tests / catalog drift checks, but the matching REA path requires external native debuggers/decompilers not installed on the CI runner:

Area Why Unverified Here
Ghidra / native decompilation depth Requires a local Ghidra install + REA bridge
Frida / LLDB / GDB / Rizin live sessions Requires those debuggers + device/process attach
JADX / full Android APK pipeline APK resigning unsupported; standalone DEX shield is in-tree
Firmware containers Staging/extraction budgets exist; no honest E2E without firmware corpora
SPIR-V / CUDA fatbinary / AMD GPU objects Need native GPU object decoders; PTX text is shielded in-tree
EVMole in-process EVM decode Not pathable in-process; EVM shield densifies to the 4 MiB input cap
REA's own PCAP provider (inspect_pcap_capture) REA runs it only on Linux / Windows; on macOS the harness runs the same two tshark passes and REA's parser
Raised REA budgets, other than JavaScript strings Not measured. Every other row is REA at its default settings

🛡️ Format Coverage

All 26 format shields are built directly into bla-agents with zero external packer dependencies:

Format Canary + Diagnostic Exhaustion / Budget Trip Empirically vs. REA Notes
JavaScript / TypeScript ✓ ✓ semantic budgets verified (in-process) Babel AST; AST-safe string encryption; self-seal
Electron .asar ✓ ✓ header 16 MiB verified IPC de-literalization + Pickle repackaging
Electron fuses ✓ ✓ >32 sentinels verified V1 fuse hardening + sentinel flood
WebAssembly (.wasm) ✓ custom sections ✓ export decoys → Wabt verified (Wabt) LEB128 parser; preserves WebAssembly.validate
SQLite (.sqlite / .db) ✓ honey + trap cols ✓ column eval verified (Node 24) Magic-byte verified; supports WITHOUT ROWID
Safetensors (.safetensors) ✓ string metadata ✓ JSON nesting >32 verified HuggingFace-loadable; tensor bytes preserved
GGUF (.gguf) ✓ ✓ ARRAY nesting >32 verified Preserves model weights and KV tensors
Go buildinfo (ELF) ✓ ✓ >1 MiB verified (format rejection) Mapped ELF section
Windows PE (.exe / .dll) ✓ ✓ .enclave COFF section + fuses/rustc/Dart verified Mapped .enclave section, x86-64 ABI function islands, MSVC RTTI, CodeView RSDS PDB
Flutter / Dart pool ✓ ✓ URI/token caps verified (projected) Snapshot magic & toolchain flood
Erlang BEAM (.beam) ✓ ✓ chunks / atoms verified Exceeds 64-chunk budget
EVM + Solidity meta ✓ ✓ → 4 MiB input verified (approach) CBOR metadata + bytecode densification
PCAP (.pcap) ✓ ✓ REA's tshark field pass > 8 MiB (fixed) verified (tshark 4.4.1) Synthetic packet + canary injection; a DNS answer flood at bankrupt
Python (.py / .pyc) ✓ ✓ designed-for CPython 3.7–3.14 marshal & AST support
NVIDIA PTX (.ptx) ✓ entry flood benchmark only Kernel entry & canary flood
DEX / IL2CPP / bplist / .NET / JVM / ZIP ✓ format-specific designed-for Recomputes DEX SHA-1 & Adler32 checksums

📊 Benchmark (14 Targets, Internal Estimates)

Run npx bla-agents benchmark to execute the 14-target showcase locally:

npx bla-agents benchmark

Note: Token and USD columns below are internal estimates derived from BLA's catalog of REA limits to illustrate relative amplification across formats — for actual process-level REA errors, see the Empirical Results table above.

Aggregate Showcase Result: ~1,068.7× token amplification · ~503.5× cost amplification · 14/14 bundled showcase targets passed runtime-equivalence checks. This result is limited to the benchmark fixtures.

Target Artifact Unprotected Baseline (est.) BLA Protected (est.) Token Multiplier Runtime Equivalence
DX-Ball sound-pan (.js) 84/100 · 4,800 tok · $0.20 2/100 · 15.2M tok · $238.02 · LIMITED 3,172.7× ✅ 100%
Notion Electron clipboard (.asar) 98/100 · 9,132 tok · $0.34 2/100 · 23.1M tok · $358.57 · LIMITED 2,532.3× ✅ 100%
Enterprise license validator (.ts) 87/100 · 4,800 tok · $0.20 2/100 · 15.2M tok · $238.02 · LIMITED 3,172.7× ✅ 100%
NVIDIA PTX compute kernel (.ptx) 81/100 · 4,800 tok · $0.20 7/100 · 23.1M tok · $358.20 · LIMITED 4,814× ✅ 100%
Erlang BEAM module (.beam) 81/100 · 8,900 tok · $0.34 5/100 · 3.86M tok · $62.08 · LIMITED 433.2× ✅ 100%
WASM crypto kernel (.wasm) 81/100 · 4,800 tok · $0.20 6/100 · 1.83M tok · $30.23 · LIMITED 381.5× ✅ 100%
Edge AI Safetensors (.safetensors) 81/100 · 4,800 tok · $0.20 6/100 · 1.81M tok · $29.90 · LIMITED 376.8× ✅ 100%
libpcap network capture (.pcap) 81/100 · 4,800 tok · $0.20 7/100 · 23.1M tok · $359.27 · LIMITED 4,817.7× ✅ 100%
Android Dalvik bytecode (.dex) 81/100 · 8,900 tok · $0.34 7/100 · 13.4M tok · $209.39 · LIMITED 1,510.5× ✅ 100%
EVM Solidity metadata (.bin) 81/100 · 6,440 tok · $0.26 6/100 · 1.78M tok · $29.40 · LIMITED 276.3× ✅ 100%
.NET managed assembly (.dll) 81/100 · 8,900 tok · $0.34 6/100 · 2.01M tok · $33.22 · LIMITED 226.3× ✅ 100%
Apple binary plist (.plist) 81/100 · 6,440 tok · $0.26 8/100 · 1.79M tok · $29.54 · LIMITED 277.6× ✅ 100%
SQLite entitlement DB (.sqlite) 81/100 · 10,568 tok · $0.29 7/100 · 1.71M tok · $28.39 · LIMITED 162.1× ✅ 100%
Native ELF64 (Rust/Go/Fuse/Dart) 81/100 · 35,222 tok · $0.73 2/100 · 3.72M tok · $60.03 · LIMITED 105.6× ✅ 100%

🎯 Red-vs-Blue CrackMe Arena

Test your reverse-engineering agent against public BLA-hardened challenges (SEAL.json publishes SHA-256 digests only — no plaintext keys or flags in git):

Challenge Architecture Hardening Highlights Try It Locally
CrackMe v6 Node Vault + HTML Panel + WASM Core bankrupt intensity, self-sealed sibling hashes, WASM export obfuscation npm run crackme:v6 -- '<key>'
CrackMe v7 Polyglot Vault (JS + WASM + Rust + Go + C) Cross-language state entanglement & multi-binary integrity gates npm run crackme:v7 -- '<key>'
CrackMe v8 Aegis Split-Key Vault AST-encrypted strings (bla_encrypt_strings) + caller scope binding (bla_bind_scope) npm run crackme:v8 -- '<key>'
CrackMe v9 Agent-Woven Entangled Vault Full BLA 3.x closed-loop weave (weaveScore >= 70) + distributed anti-strip checks npm run crackme:v9 -- '<key>'

🔒 Security & Configuration Defaults

  • Per-Project Random Seed — BLA automatically generates and persists a unique cryptographic seed in .blarc / bla.config.json on first run so two projects never share decoy identifiers.
  • Strict Packaging Lint Gates (bla lint) — Refuses bankrupt hardening when plaintext FLAG{...} secrets, single-point integrity checks, or unbound exported verifiers remain in source.
  • Opt-In --llm-directives — Aggressive in-band prompt decoys are off by default; pass --llm-directives (or llmDirectives: true) to enable.
  • Opt-In Signed Attestation (--manifest) — Emit .bla-manifest.json with --manifest, sign with --sign-manifest --manifest-key <pem>, and verify both ed25519 signature and artifact SHA-256 via bla verify-manifest <file> --public-key <pem>.

❓ FAQ

Which coding agents can use BLA?

Any coding agent supporting MCP or SKILL.md workflows. npx bla-agents setup auto-registers Claude Code, Cursor, Codex, Gemini CLI, Windsurf, Google Antigravity, OpenCode, VS Code, and GitHub Copilot CLI. After setup, the bundled blind-llm-agents skill guides your agent through the closed-loop workflow (bla_weave_plan → source weave → bla_harden → bla_verify).

Does BLA slow down my app or change runtime output?

No on the hot path. Exhaustion graphs sit behind O(1) number-theoretic opaque predicates, and binary/model/database decoys live in inert metadata or honey tables. Every benchmark target verifies 100% input/output equivalence against the original build.

Do I need Ghidra, Hopper, IDA, or Wabt installed to protect files?

No. All 26 format shields are built into bla-agents in pure TypeScript/Node.js. External tools (Wabt, tshark) are only used if you run the optional live REA red-team suite (npm run redteam:rea).

Can BLA be used to attack other people's software?

No. BLA is strictly a Blue Team defensive tool that hardens your own build artifacts before release. See DISCLAIMER.md.


🤝 Contributing & Community

Issues and PRs are welcome at qybaihe/bla. Before submitting a PR:

npm run build && npm test && node dist/cli.js benchmark
node scripts/verify-website.mjs

📄 License

MIT

Star history

Star History draws the GitHub star growth for qybaihe/bla. The chart image is cached for about a day and is not a per-second count. Select it to open the full history.

Star history chart for the qybaihe/bla GitHub repository
S
Description
🛡️ BLA(盲式语言模型代理)——一种对抗性防御护盾,能够保护二进制文件、Electron 应用程序、JavaScript 模块、Python 代码以及 WebAssembly 代码免受像 REA 这样的 AI 逆向工程代理的攻击。毒化 -> 释放 -> 假象生成。
Readme MIT 13 MiB
Languages
TypeScript 59.7%
JavaScript 31.3%
HTML 7.9%
Python 0.6%
CSS 0.5%