Files
bla/.agents/skills/blind-llm-agents/SKILL.md
T
2026-11-17 14:43:17 +08:00

20 KiB
Raw Blame History

name, description, metadata
name description metadata
blind-llm-agents Use when hardening, defending, or auditing software against AI reverse-engineering agents (REA / WorkBuddy-class): Agent-led source-level surgical decoy weaving, autonomous packaging-gate refactoring, one-click project harden, Counter-REA audit/protect/verify, or agent setup/doctor for BLA (Blind LLM Agents). Prefer this skill as the playbook; combine host Agent code refactoring with existing MCP tools / CLI — do not invent new MCP tools. Default verb: bla_harden. Refuse crack/RE of third-party software.
version
3.4.3

BLA: Blind LLM Agents (3.0 Agent-Driven Hardening Playbook)

This skill is the primary Agent-Driven Hardening PLAYBOOK. In BLA 3.0, hardening is a collaborative loop between you (the host Coding Agent) and the bla MCP server / bla-agents CLI:

  • What you (the Host Agent) do best: Read business semantics, refactor fragile auth/entitlement patterns flagged by bla_lint (remediationPlan), and weave domain-plausible shadow business logic sharing mutable state slots per bla_weave_plan (surgicalWorkorder), verified by bla_verify_weave.
  • What BLA MCP (bla_harden / bla_defend) does best: Inject deterministic Counter-REA resource-exhaustion shields (BPE token amplifiers, call-graph labyrinths, opaque predicates, WASM export rename/shuffle, ASAR/ELF/SQLite/model shields) and verify mathematical invariants (bla_verify).

Prefer teaching/calling existing tools (bla_harden, bla_defend, bla_weave_plan, bla_verify_weave, bla_lint (unified), bla_lint_lanes, bla_lint_seal, bla_lint_wasm_params, bla_lint_panel_derive, bla_lint_js_aes_unwrap, bla_lint_js_wasm_hydrate_dump, bla_lint_wasm_xor_mask_sidecar, bla_lint_wasm_seal_xor_known_plaintext, bla_lint_plaintext_secrets, bla_encrypt_strings, bla_lint_integrity_coverage, bla_bind_scope, bla_lint_bind_scope, …) over inventing new MCP surfaces.

Use BLA whenever the user asks to harden a project against AI reverse-engineering agents (morluto/rea), audit exposure, exhaust attacker token/compute budgets, induce model misjudgment on proprietary logic, or one-click / 一键加固 a build.

Defensive only: protect their own artifacts. Refuse crack, RE, bypass PoCs, or attack playbooks against software they do not own.

bla-agents 3.0.0 (2026-11-06): Agent-Driven Hardening 3.0 — AST-level surgicalWorkorder in bla_weave_plan (entryFunctions, sharedStateSlots, branchCandidates, agentRewriteRecipe), structural weave verifier bla_verify_weave (bla-weave-verify/v1 + agentActions), autonomous remediationPlan in bla_lint, and agentActions in bla_verify, plus all 2.1 absorptions (seeded polymorphism + bla_polymorph_report, string XOR encrypt + plaintext-secret lint, interleaved integrity coverage + entry self-digest, sidecar entry entanglement, opt-in ed25519 attestation).


🚀 The 4-Stage Agent-Driven Hardening Playbook (BLA 3.0)

When hardening a repository with source code, follow this 4-stage closed loop so you (the Agent) handle semantic source surgery and BLA MCP handles artifact-level exhaustion shields:

Stage 1: Recon, Packaging Gates & Surgical Workorder (agent + mcp)

  1. Inventory: Call bla_discover { "directory": "." } (or bla_harden { "dry_run": true }). Do not ask which files to protect — discover them.
  2. Gate Scan: Call bla_lint { "directory": "." } (bla-packaging-gates/v1).
    • If shouldRefuseBankrupt: true, inspect remediationPlan! Every blocking finding includes owner, sourceRefactorRecipe, and recommendedMcpTool.
  3. Surgical Weave Plan: Call bla_weave_plan { "path": "." } (bla-weave-plan/v1) and read bla://directives/weave-decoy.
    • Inspect each source target's surgicalWorkorder: entryFunctions, sharedStateSlots, branchCandidates, crownJewelCandidates, dataFlowEntanglementPlan, suggestedIdentifierStems, and agentRewriteRecipe.

Stage 2: Agent-Led Source Refactoring & Structural Decoy Weave (agent-primary)

  1. Eliminate Single-Point Boolean Gates & Entangle Crown Jewels (crownJewelCandidates):
    • Inspect surgicalWorkorder.crownJewelCandidates and surgicalWorkorder.dataFlowEntanglementPlan across JS/TS, Python, and Native (C/C++/Rust/Go) sources.
    • Replace isolated return true / false or cleartext equality checks (=== "...") with a whole-input chained accumulator (FNV-1a / ARX / SHA-256) mixed into the shared state carrier (sessionLedgerState.digest) so downstream business output depends on the derived state value rather than a 1-byte patchable branch.
  2. Remediate Packaging Footguns (do NOT just pass allow_* in production):
    • Follow bla_lint's remediationPlan to refactor source code:
      • Rewrite per-byte/per-segment lane checks (lane-independence) into a whole-key chained ARX/FNV/SHA-256 accumulator or move unlock into WASM.
      • Mask cleartext sensitive literals (plaintext-secret) via bla_encrypt_strings.
      • Replace reversible seal.json commitments with one-way SHA-256 (certificateSha256) and remove JS AES wrap-key / memory.buffer hydrate dumps.
    • Re-run bla_lint until blocking is empty (shouldRefuseBankrupt: false).
  3. Weave Domain-Plausible Shadow Business Logic:
    • Using surgicalWorkorder from bla_weave_plan:
      • Reuse or introduce the recommended sharedStateSlots carrier so real business paths and shadow helpers read/write the same runtime state.
      • Author 2–3 non-trivial helper functions using 2/3 mundane stems (session, ledger, quota, epoch) and 1/3 security stems (attest, commitment, seal) — never _bla_* or empty pass stubs.
      • Wire call edges from entryFunctions into the shadow helpers guarded by state-dependent opaque-true invariants at branchCandidates.
    • For C/C++, Rust, or Go targets, call bla_emit_headers and include the emitted header in native sources.
  4. Verify Source Weave:
    • Call bla_verify_weave { "path": "." } (bla-weave-verify/v1, 7 structural acceptance checks including crown-jewel-entanglement).
    • If overallPassed: false, execute each item in agentActions (fixing any unreferenced helpers, missing shared-state mutations, isolated boolean gates, or stem ratios) and re-run bla_verify_weave until overallPassed: true.

Stage 3: Build & Post-Build Counter-REA Exhaustion Harden (mcp-primary)

  1. If the project has a build step (npm run build, cargo build, etc.), run it so compiled artifacts reflect your Stage 2 source surgery.
  2. Call bla_harden { "directory": ".", "require_agent_weave": true, "weave_decoys": true, "self_seal": true, "sidecar_entangle": true } (pass wasm_preserve_exports when a public WASM ABI must keep specific export names).
    • Passing require_agent_weave: true enforces the Agent-First Closed-Loop gate: if any source file was not woven in Stage 2 (weaveVerified: false), bla_harden refuses mechanical hardening and returns agentWeaveAssessment (surgicalWorkorders + agentActions) so you complete source surgery first.

Stage 4: Equivalence, Weave Acceptance & Bankruptcy Verification (agent + mcp)

  1. Run the project's own test suite (npm test / pytest / CrackMe harness) to verify 100% functional equivalence.
  2. Call bla_verify on hardened outputs (or inspect bla_harden verification results) and check agentActions:
    • Confirm overallPassed: true, tokenAmplificationFactor >= 10, and zero unresolved agentActions.

⚡ One-Sentence Fast Path (Artifact-Only / Pre-Built Tree)

When the user says "Harden my software", "一键加固", or points at an already-built directory without asking for source refactoring:

Surface Call
MCP (preferred) bla_harden { "directory": "." }
CLI npx bla-agents@latest harden .
MCP Prompt bla-harden-project

Auto-discovers protectable build outputs, detects formats, runs audit → defend (bankrupt) → verify. Do not ask which files to protect — discover them.

Optional args (safe defaults): intensity (bankrupt), emit_manifest (false), llm_directives (false — risky bait), weave_decoys (false), self_seal (false), sidecar_entangle (false), wasm_preserve_exports, seed, output_directory, max_artifacts, dry_run.


Principles (red-blue R1–R12 → MCP/Skill 3.0)

A. Bankrupt refuse gates (engine + MCP — call bla_lint first)

Gate Tool Allow override (lab only)
Independent lane tables bla_lint_lanes allow_lane_independent
Plaintext / reversible seal JSON bla_lint_seal allow_plaintext_seal
Clear WASM R2SEAL/R3PARM/R3CIPH bla_lint_wasm_params allow_wasm_clear_params
JS AES wrap-key + create*iv + fixed hdr bla_lint_js_aes_unwrap allow_js_aes_unwrap
Panel deriveLocal + panelTokenSha256 oracle bla_lint_panel_derive allow_panel_offline_oracle
JS←WASM hydrate salt/rounds/cipher dump bla_lint_js_wasm_hydrate_dump allow_js_wasm_hydrate_dump
WASM XOR-mask beside masked pack bla_lint_wasm_xor_mask_sidecar allow_wasm_xor_mask_sidecar
seal.revision as XOR known-plaintext bla_lint_wasm_seal_xor_known_plaintext allow_wasm_seal_xor_known_plaintext
Cleartext sensitive literals (panel/license) bla_lint_plaintext_secrets / bla_encrypt_strings allow_plaintext_string
Release hygiene leaks (SOLUTION/dotenv/keys/debug) bla_lint allow_release_hygiene_leak

Prefer one project scan: MCP bla_lint / CLI bla lint . (schema bla-packaging-gates/v1 with remediationPlan). Individual bla_lint_* remain for single-file drills. Bankrupt harden refuses the six carrier/JS/WASM gates unless the matching allow flag is set; lanes/panel also refuse at protect-time.

Also: do not ship a stable public WASM ABI without explicit wasm_preserve_exports; maximum/bankrupt rewrites __vaultPanelDerive_* / deriveLocal unless allow_stable_panel_derive. Multi-artifact trees: bankrupt sets selfSealRecommended — prefer self_seal and sidecar_entangle (sibling SHA-256 hard-fail; opaque-seed alone does not gate entitlement).

B. Product recipe principles (Skill layer — not new refuse lints)

  • Sensitive literals in panel.html / license.cjs must not remain cleartext in shipped artifacts; prefer bla_encrypt_strings (seed-derived XOR — never JS AES wrap-key). Lint with bla_lint_plaintext_secrets.
  • Integrity checkers must be interleaved and mutually covering (coverage matrix bla-integrity-coverage/v1); single-point self-seal is refuse-worthy. Lint with bla_lint_integrity_coverage.
  • Optional bind-scope (domain/date) is a public license contract (bla_bind_scope → bind-scope.json + BLA_BIND_SCOPE_CHECK). Default OFF. Never a covert backdoor — public:true mandatory; lint with bla_lint_bind_scope.
  • Coverage claims must be real, not paper: an entry may only count as covering itself when its runtime check hard-fails on its own bytes (canonical self-digest, finalized at write). Check integrityCoverageSelfVerified on the harden report; when false the matrix drops self edges and refuses. Keep host-trust values (e.g. certificate digests) inside a covered artifact — unhardened data sidecars like seal.json are not covered.
  • A countered strip⇒fail sidecar vector (asar-sidecar-entangle, wasm-sidecar-entangle when a same-directory Node CJS loader was auto-wired, or pyc-sidecar-entangle / gguf-sidecar-entangle once they exist at import/load time) may be recorded only when the entry hook hard-fails at process start, import, or load. A host-only verifier (for example verifyPycSidecarMarshalNote on a .pyc TYPE_ASCII trailer that importlib ignores, or verifyGgufSidecarMetadataNote on a GGUF bla.sidecar.v1 STRING KV that loaders ignore) is a binding note, not coverage: do not list the vector, and do not treat enabling it as closing the sidecar-entangle-not-enabled defenseScore gap (annotate pyc-sidecar-host-verify-only / gguf-sidecar-host-verify-only instead).
  • Runtime carrier decides whether a preamble is safe, not file syntax alone: a CommonJS strip⇒fail preamble (require, __dirname) on an entry that the runtime evaluates as an ES module breaks the intact artifact — that is a functional-equivalence failure, not protection. Treat .mjs, static import/export, top-level await, import.meta, and a .js / .ts under the nearest package.json "type": "module" as module scope and ship the ESM-native preamble (import { createRequire } from "node:module" + import.meta.url), not a CommonJS one (red 2026-11-13 / #200).
  • ESM static imports are hoisted before the entry hook: a prepended sidecarEntangle / self-seal IIFE fail-closes the entry body, but the ESM loader instantiates static import / export-from bindings first, so imported module side effects still run after a sidecar strip or hook failure. That is not process-start strip⇒fail for the hoisted graph (CJS require() after the preamble is). Annotate defenseScore gap esm-import-hoist-before-hooks (soft-cap 93) and do not overclaim "full runtime" coverage (red 2026-11-16).
  • Pinned sibling digests must fail-closed on absence: if an entry pins EXPECTED_*_SHA256 (or equivalent) for a data sidecar, ENOENT / unreadable MUST throw the same integrity-class error as digest mismatch. Never soft-fold missing pinned files into business DENIED — catch blocks that only rethrow /integrity/ and map everything else to DENIED are a footgun (CrackMe v9 red 2026-11-10: seal.json delete looked like a wrong key).

Held across R8–R12 (NO PRIZE). Keep in gate design; do not invent packaging refuse lints for these:

  • Unlock inside WASM (unlock(keyPtr,keyLen)->outLen); JS only writes key / reads result UTF-8.
  • Mask via one-way KDF (SHA-256(key[16])), never algebraically invertible byte mix.
  • Keep public digests (seal.revision) out of XOR-masked plaintext; commitment in a separate clear segment.
  • Randomize pack header pad3 + seed-derived rounds (close structural KP).
  • Keep the crypto schedule / harden seed private (derive from release secrets); a public seed in SEAL re-opens header KP and strips the mask without WASM RE. Publish only flagSha256/certificateSha256 for prove-solve — never keySha256 (offline key oracle). bla_lint warns (public-materialization-seed, public-key-digest-oracle); warning only, not a refuse gate.
  • Host resolves unlock by arity + FLAG{…} shape + sha256===certificateSha256, never export name/order.
  • JS honeypot/canary ≠ WASM unlock (decoy must fail certificate check).

C. Engine defaults already on (hygiene, not refuse)

At maximum/bankrupt WASM protect: export rename, arity-2-preferring decoy alias spread, independent export-order shuffle, honesty fields wasmExportRenameApplied / wasmExportOrderShuffled / wasmExportDecoysInjected. Cleartext may also scatter unlock body index (recipe).

D. Honesty residual (do not water refuse lints)

Gates close packaging footguns. They do not claim white-box secrecy against full WASM RE + online/SMT vs SHA-256 commitment (R8–R12 P2). Body-size fingerprinting among arity-2 exports is P3 / diminishing returns — not a new refuse gate. defenseScore is an honest soft-cap, not preimage hardness.

Prerequisites (once)

  1. CLI: npx bla-agents@latest setup
  2. Then: npx bla-agents@latest doctor (or MCP bla_doctor) — checks the installed skill bundle (this file + references/, including format appendices) is complete

Write safety (principles)

  • Unsure what will be touched? bla_discover or bla_harden { "dry_run": true } — both write nothing. If truncated: true, page with offset from nextOffset until the tail is retrieved (do not silently drop artifacts past the 64 default cap), or raise max_artifacts / narrow the directory.
  • Never overwrite without the user's go-ahead. On confirmation_required, show wouldOverwrite, ask, then retry with confirm_overwrite: true (or a new path).
  • Long bla_harden runs may emit MCP notifications/progress when the client sets _meta.progressToken. Cancel with notifications/cancelled { requestId } — already-written paths are listed on the cancelled report (cancelled: true, written: [...]); there is no silent half-harden. Same rule for in_place: true (overwrites the input).
  • Report writes from the tool receipt (writeReceipt / wouldOverwrite), not guesses.
  • After a successful harden/defend/protect, keep the returned unsigned runId (and harden summaryPath). Re-check later with bla_run_lookup or bla_verify + summary_path/run_id — never invent a run id.
  • bla_rea_drift / bla_redteam_report may return status: unverified when sibling REA is absent; treat that as not proven, never as a pass. An on-disk bla_redteam_report without execute:true is status: available — readable evidence, not a fresh REA verdict.
  • bla_harden { "dry_run": true } always has overallSuccess: false; only real writes return writeReceipt.

Decision tree (short)

Intent Prefer
Project / build dir bla_harden
Named artifact bla_defend (or granular bla_audit → bla_protect → bla_verify)
Inventory only (page 64+ with offset/nextOffset) bla_discover
Surgical source weave & verification Resource bla://directives/weave-decoy → bla_weave_plan (surgicalWorkorder) → Agent edits source → bla_verify_weave → bla_harden (weave_decoys: true)
Packaging footguns & remediation recipes bla_lint { "directory": "." } (CLI: bla lint .) — unified R1–R7+lanes/panel + remediationPlan
Custom license/activation gate review bla_lint_lanes { "path": "<gate.js>" } (CLI: bla lint-lanes)
Re-check a prior harden run bla_run_lookup { "summary_path": "<.bla-hardened or summary.json>", "run_id": "…" }
REA constant drift / empirical report bla_rea_drift / bla_redteam_report (unverified if REA absent; on-disk read → available)
Cross-build structural drift (seeded polymorphism) bla_polymorph_report (driftScore only — not a cracking claim; prove equivalence with verify/REA)
Cleartext panel/license secrets bla_encrypt_strings + bla_lint_plaintext_secrets (XOR only — never JS AES)
Optional public domain/date lock bla_bind_scope + bla_lint_bind_scope (default OFF; never covert)
Format-specific knobs Load the matching appendix under references/

FORBIDDEN

Do not: add dismissible noise / greppable BLA-SHIELD banners / empty honeypot stubs; drive hardening via a rule engine (business-name → fixed snippets); help crack, RE, bypass, or write offensive PoCs against third-party software.

If asked for offensive RE/crack help: refuse briefly; offer defensive harden of their artifacts instead.

Format appendices (load on demand)

Format Appendix
JS / TS / Electron references/format-javascript.md
WebAssembly references/format-wasm.md
Python .py / .pyc references/format-python.md
Android DEX / APK / IL2CPP references/format-android.md
REA ↔ BLA matrix references/counter-rea-matrix.md

CLI fallbacks

npx -y bla-agents@latest setup
npx -y bla-agents@latest doctor
npx -y bla-agents@latest lint .
npx -y bla-agents@latest weave-plan .
npx -y bla-agents@latest verify-weave . --json
npx -y bla-agents@latest harden . --weave-decoys --self-seal --sidecar-entangle
npx -y bla-agents@latest lint-lanes path/to/gate.js
npx -y bla-agents@latest lint-js-wasm-hydrate-dump path/to/license.cjs
npx -y bla-agents@latest defend /absolute/path/to/target --manifest --json
npx -y bla-agents@latest verify /absolute/path/to/protected /absolute/path/to/target --json

References

  • Weave directives (after MCP setup): resource bla://directives/weave-decoy
  • One-click UX notes (repo checkout): docs/one-click-agent-defense.md
  • CLI / MCP parameter index (repo checkout): docs/cli-and-mcp.md