Files
bla/STATE.md
T
qybaihe 0edd5f8deb docs: sync STATE/skill/CHANGELOG for wasmPreserveExports and #40 progress
Bump blind-llm-agents skill to 1.3.2; mark landed #40 items in STATE and
weave-decoy follow-ups.
2026-11-02 21:39:38 +08:00

7.4 KiB
Raw Blame History

BLA state of the project

Version: 1.6.4 (package.json ↔ src/identity.ts, verified by build; see CHANGELOG.md)
Status: CI matrix runs typecheck, build, tests and the benchmark on ubuntu / macos / windows with Node 24. Windows checkouts keep challenge/** LF via .gitattributes so the CrackMe seals verify.
Publishing: tag-only releases (git tag bla-agents-<version>); see CHANGELOG.md and GitHub Releases.

One-sentence usage

npx bla-agents setup -y    # register MCP + skill (default verb: bla_harden)
bla harden .               # or tell your agent: "Harden my software against REA"

Also: bla_harden MCP tool with { "directory": "." }.

What BLA covers

Format Canary + diagnostic + exhaustion Empirical vs real REA Notes
JavaScript / TypeScript ✓ verified — real REA resourceLimits (alias/getter/primitive) + anti-strip AST + semantic budgets
Electron ASAR ✓ verified — DEFAULT_ASAR_MAX_HEADER_BYTES
WebAssembly ✓ export decoys verified — Wabt WASM_INTERFACE_REPRESENTATION_BYTES / partial needs Wabt for REA path
Python .py / .pyc ✓ designed-for .py AST weave + .pyc remmarshal co_consts CODE weave when weave_decoys
Native ELF (Go/Flutter/Electron fuses) ✓ verified — MAX_BUILD_INFO_BYTES, pool URIs/tokens, fuse sentinels
SQLite ✓ honey + trap columns verified on Node 24 — SqliteColumnEvaluationFailure; Node 22 soft capability_gated
Safetensors ✓ nesting breaker verified — MODEL_JSON_NESTING (HF-loadable)
GGUF ✓ verified — MODEL_METADATA_NESTING
Erlang BEAM ✓ verified — max_chunks
EVM bytecode ✓ verified — EVM_INTERFACE_LIMITS.inputBytes
PCAP ✓ verified — TSHARK_JSON_MAX_BYTES (needs tshark)
GPU PTX ✓ canary/exhaustion soft — REA still after=ok densifies but no hard budget trip
DEX / IL2CPP / bplist / .NET / JVM / ZIP / Android R8+JNI ✓ designed-for (R8/JNI new in 1.4.6) signed APK refused (by design); unsigned APK gets mapping.txt + decoy JNI .so
Firmware UF2 ✓ canary/exhaustion designed-for — counters REA #8230 (not-main-flash / file-container decoys) added in 1.6.2
SPIR-V / CUDA / AMD GPU objects — skipped no stable REA analyzer / not in scope yet

Designed-for but not empirically verified here (need native tooling on the REA side): Ghidra, Frida, LLDB, Rizin, JADX.

Empirical red-team headlines (npm run redteam:rea)

Commodity hardware, sibling REA, Node 24.22.1 + Wabt 1.0.42 + tshark:

Format Before → after REA budget / error
Go buildinfo 51 339 → 1 214 706 B MAX_BUILD_INFO_BYTES
Flutter pool same ELF plant MAX_POOL_URIS / MAX_POOL_TOKENS
BEAM 84 → 267 876 B max_chunks
EVM 16 → 4 190 208 B EVM_INTERFACE_LIMITS.inputBytes
PCAP 24 → 1 348 012 B TSHARK_JSON_MAX_BYTES
Safetensors 128 → 10 328 B MODEL_JSON_NESTING
GGUF 24 → 7 840 B MODEL_METADATA_NESTING
ASAR 607 → 17 752 989 B DEFAULT_ASAR_MAX_HEADER_BYTES
WASM 54 → 6 878 763 B WASM_INTERFACE_REPRESENTATION_BYTES
SQLite (N24) 12 288 → 57 344 B SqliteColumnEvaluationFailure
Electron fuses same ELF plant maximumSentinelsPerSlice
JavaScript 761 → 116 210 B alias-selections+getter-expansion+primitive-bytes+primitive-candidates
GPU PTX 91 → 26 424 B still soft (after=ok)

npm run verify:rea — catalog drift harness only (42 checks). Not protection proof. npm run redteam:rea — only authoritative evidence (real in-process REA exceptions).

Benchmark (internal estimates)

14-target Counter-REA showcase: ~612.5× token amplification, ~292.2× cost, 100% runtime-equivalent. Numbers are internal model estimates, not empirical REA process measurements. Do not cite as red-team proof.

Dogfood (local user journey)

Outside the repo, npm pack → install tarball → isolated HOME:

  1. npx bla-agents setup -y --client cursor → APPLIED; skill default verb bla_harden
  2. bla harden . → discovered 10 artifacts (dist/* + root tools.py / local.sqlite), hardened 10/10, doctor HEALTHY
  3. Hardened JS/WASM/SQLite/Safetensors/ELF/ASAR/Python remain valid and runtime-equivalent
  4. MCP tools/list includes bla_harden; tools/call bla_harden succeeds

Done this critique pass

  • Anti-strip / realistic identifiers + CF entanglement (JS AST)
  • Orders-of-magnitude tarpit scaling (not keyed to exact REA consts)
  • redteam:rea as sole authoritative evidence; verify:rea demoted
  • Honest docs separation verified / designed-for / skipped

Nightly progress (2026-10-28 Asia/Shanghai)

  • Pulled both repos. REA delta since last local tip: 325 commits on main (dd7a2da5b → 0cd7ef456). Headline new analyzers: Android R8/ProGuard mapping (#7456), JNI native-declaration ↔ .so export linking (#7472), manifest component declarations (#7379). Most other delta is evidence/contract refactors.
  • Shipped v1.4.6 counter: android-r8.ts + unsigned-APK wiring + identity constants + verify:rea drift entries.

Nightly roadmap remaining

  1. Self-seal / defenseScore honesty (#40) — landed soft-caps + opt-in --self-seal / MCP self_seal / .blarc selfSeal (batch 2026-11-02). Remaining: deeper onefile/peel score model.
  2. Sidecar entry entanglement — DESIGN + planSidecarEntanglement() stub landed; full strip⇒fail rewrite still open.
  3. Deception shields → more formats — WASM export rename + wasmPreserveExports landed; DEX/SQLite/model densify and broader numeric template hygiene still open.
  4. Empirical JADX / Android red-team — real REA mapping+JNI vs BLA-hardened unsigned APK.
  5. ESP flash / esp-image (#8228) — if REA moved there; UF2 empirical red-team otherwise.
  6. Dependabot majors #29–#35 — still held pending review.
  7. Polish doctor subsystem list (SQLite / models / BEAM / EVM / Android R8) for parity with shields.
  8. Lane-independence lint (#40.1) — bla lint-lanes heuristic shipped; optional CI gate / doctor hook TBD.

Publish checklist (once a GitHub token exists)

  1. Confirm package.json version === src/identity.ts === CHANGELOG / docs/releases/v1.4.0.md (already 1.4.0)
  2. Final gate matrix on Node 22.23.4 and Node 24.22.1 (+ Wabt for redteam):
    npm run build && npm test && npm run benchmark && npm run verify:rea && npm run redteam:rea
  3. Prefer Node 22.23.4 / 24.x for REA (avoid ~/node22 22.13.0 — lacks zstdDecompressSync)
  4. git push origin main (local commits are ahead; remotes currently untouched)
  5. git tag v1.4.0 && git push origin v1.4.0
  6. npm publish for bla-agents (and alias/bla-shield dual-pack flow if still used)
  7. GitHub Release from docs/releases/v1.4.0.md
  8. Smoke: npx bla-agents@latest setup -y + bla harden . on a fresh machine

Honesty red line

BLA is a defensive tool for protecting your own software. See DISCLAIMER.md. Do not overclaim: separate empirically verified commodity hits from designed-for / unverified native-tool paths and skipped formats.