Skip to content

armgit / bla

Public Imported from GitHub
0

🛡️ BLA (Blind LLM Agents) — Adversarial defense shield that protects binaries, Electron apps, JS bundles, Python, and WebAssembly against AI reverse-engineering agents like REA. Poison -> Exhaust -> Hallucinate.

3.2 KB
Raw Blame

Counter-REA Model-Misjudgment Induction (src/shields/deception/)

While traditional obfuscators only try to hide code, BLA actively induces LLM reverse-engineering agents to reconstruct the wrong algorithm by exploiting REA's Evidence Contract.

Exploiting REA's COMMON_DISCIPLINE (promptContracts.ts:52-57)

REA instructs coding agents:

  1. "Report Observations only from cited Evidence returned by tools. Label reasoning as Inference with confidence and competing explanations."
  2. "Keep incomplete coverage, unsupported capabilities, and conflicting evidence explicit as Unknowns."

BLA exploits this discipline through a two-pronged pincer movement:

  • Prong A (Demote Real Logic to Unknown): Wrap the real business logic in REA Semantic Budget Breakers (primitive-candidates > 256, alias-selections > 512) so REA's static analyzer tags the real logic as status: "unknown".
  • Prong B (Promote Shadow Logic to High-Confidence Observation): Inject crystal-clear, syntactically pristine Shadow Business Logic (shadow-business-logic.ts), Semantic Honeypots (semantic-honeypot.ts), Evidence Contract Competing-Explanation Lures (evidence-contract-poison.ts), and Multi-Tier gRPC/vsock Architecture Mirages (architecture-mirage.ts).

When the LLM agent queries REA's search_procedures, search_strings, and analyze_function, the Shadow Business Logic returns 100% complete observed Evidence—causing the LLM to faithfully rebuild the decoy algorithm!

Python / PyInstaller honesty

BLA's Python shield densifies .py / .pyc artifacts (interconnected tarpit, opaque predicates, diversified chaff). It does not rewrite a packed PyInstaller onefile CArchive in place: detectNativeEofAppendHazard only refuses EOF trailers that would break the MEI cookie.

Expected attacker move: peel onefile → extract entry pyc / scripts. Defense that still helps: run bla harden on sources before pyinstaller, so the peeled archive still contains dense decoys rather than clean customer logic. Opt-in self-seal / anti-debug helpers that bind the customer entry to intact bytes remain a separate product gap (issue #40).

Weave decoys (not dismissible noise)

Empty / homogeneous scare decoys are filtered as a separate layer by WorkBuddy-class agents. Prefer weaving decoys into real customer paths (shared state, entry reachability, mixed vocabulary, opaque-true happy path). JS AST already weaves; Python .py with weave_decoys:true structurally weaves decoys into customer defs (warm-init-only when the flag is false). Executable theory + MCP: weave-decoy-mcp.md, resource bla://directives/weave-decoy, tool bla_weave_plan, flag weave_decoys.

WASM export rename (#40.3)

protectWasmBinary renames exports not listed in preserveExports. At maximum/bankrupt, all non-preserved names rename by default (red-blue round-1). Keeping a stable public ABI requires an explicit preserve list. Optional renameInternalOnly: true restores the weaker internal-only pass. Returns exportsRenamed for glue updates. Project / CLI / MCP parity: .blarc wasmPreserveExports, --wasm-preserve-exports, MCP wasm_preserve_exports.