Skip to content

armgit / bla

Public Imported from GitHub
0

πŸ›‘οΈ BLA (Blind LLM Agents) β€” Adversarial defense shield that protects binaries, Electron apps, JS bundles, Python, and WebAssembly against AI reverse-engineering agents like REA. Poison -> Exhaust -> Hallucinate.

5.8 KB
Raw Blame

bla-agents v1.4.0 β€” Release notes

Publish checklist (run ONLY after a GitHub token is configured)

Do not run this sequence until gh auth status (or GITHUB_TOKEN / GH_TOKEN) succeeds. Remotes and tags are intentionally untouched in the staging commits.

# 0. Confirm auth
gh auth status
# optional if git HTTPS needs the credential helper:
gh auth setup-git

# 1. From a clean tree on main, with local commits already at 1.4.0:
cd /path/to/bla
git status   # should be clean; version 1.4.0 already committed

# 2. Push main (does not create the release tag yet)
git push origin main

# 3. Create and push the annotated release tag
git tag -a bla-agents-1.4.0 -m "bla-agents v1.4.0"
git push origin bla-agents-1.4.0

# 4. Publish the GitHub Release from this file
gh release create bla-agents-1.4.0 \
  --title "bla-agents v1.4.0" \
  --notes-file docs/releases/v1.4.0.md

# Optional: also publish npm tarballs if that pipeline is wired
# npm run release:minor   # DO NOT use β€” that would bump past 1.4.0 and push

Headline numbers

Gate Result
Tests (Node 22.23.4) 202 / 202 pass
Tests (Node 24.22.1) 202 / 202 pass
npm run verify:rea 42 drift checks (NOT protection proof)
npm run redteam:rea (Node 24 + Wabt) real-REA budget/exception hits across Go, Flutter, BEAM, EVM, PCAP, GGUF, Safetensors, ASAR, WASM, SQLite, Electron fuses and JavaScript (alias-selections+getter-expansion+primitive-bytes+primitive-candidates); GPU-PTX still soft (after=ok, not inflated)
Anti-strip gate 4 / 4 β€” hardened 790 B sample: `grep -v '_bla
14-target benchmark (internal estimate, NOT red-team proof) ~612Γ— aggregate token amplification Β· 100% runtime-equivalent on all 14
Native ELF showcase multiplier ~105.6Γ— (ISO_SIZE 48 KiB)
ASAR after harden ~17.8 MiB header past DEFAULT_ASAR_MAX_HEADER_BYTES
WASM after harden ~6.88 MiB Β· 251 exports Β· Wabt ifaceState=partial
SQLite after harden (Node 24) 57β€―344 B Β· 5 honey tables Β· SqliteColumnEvaluationFailure on trap columns
Safetensors after harden 10β€―328 B Β· MODEL_JSON_NESTING Β· HF-loadable

What's new since v1.3.0

  • Drift harness β€” verify:rea tracks REA constant movement; authoritative evidence is redteam:rea only.
  • Graceful Node degradation when node:sqlite is missing.
  • Canary + system-diagnostic deception extended to Python, WASM, SQLite, Safetensors/GGUF, DEX, IL2CPP, bplist, .NET, JVM, ZIP/APK.
  • New domains: Go buildinfo, Flutter pool, BEAM, EVM, PCAP, GPU-PTX.
  • bla_harden one-sentence UX with auto-discovery (CLI + MCP).
  • Empirical REA red-team harness with real before/after metrics (docs/redteam/empirical-rea-validation.{md,json}).
  • 14-target Counter-REA benchmark.
  • Launch-ready README with empirical-vs-estimate separation and honest coverage matrix.
  • STATE.md + dogfood polish: root-level harden discovery, CLI Quick start.

Empirical validation highlights (real REA)

Recorded under Node v24.22.1 with WABT 1.0.42 at REA_WABT_BIN_DIRECTORY (soft-skips keep Node 22 / CI green when tooling is absent):

SQLite (inspectSqliteDatabaseSnapshot)

Phase Input REA outcome
before 12β€―288 B ok β€” 1 table (license_cache), DatabaseSync.limits present
after 57β€―344 B budget_rejected β€” SqliteColumnEvaluationFailure: unknown function bla_enclave_unseal() on _bla_enclave_key_hierarchy.unsealed_preview; 5 _bla_* honey tables disclosed in schema

Honey metadata does trip REA's SQLite redaction/column-evaluation path under Node 24 (not merely the 16 MiB JSON reply budget). On Node 22 the probe soft-skips as capability_gated.

WASM Wabt E2E (WabtArtifactProvider)

Phase Input Exports Wabt interface Outcome
before 54 B 1 (computeDigest) state=complete, 1 entry, 222 B WAT ok
after 6β€―878β€―763 B 251 (250 long-name decoys + original) state=partial, 231 entries, 6β€―887β€―986 B WAT budget_rejected β€” Typed WASM interface decoding exceeds the WABT output representation budget

Original computeDigest export, WebAssembly.validate, and execute remain intact (red line held). Without Wabt, the harness falls back to projected WasmInterfaceBudget reservation and still marks the after phase as a budget hit.

Safetensors (readSafetensorsHeader / inspect_model_artifact)

Phase Input REA outcome
before 128 B ok
after 10β€―328 B capacity β€” MODEL_JSON_NESTING: Safetensors JSON nesting exceeds 32 levels

Extra bla_enclave_topology key on the first tensor descriptor; tensor bytes and required fields preserved; HuggingFace safe_open loads the file.

Other empirical budget hits (same harness)

  • Go buildinfo β†’ MAX_BUILD_INFO_BYTES
  • Flutter pool β†’ MAX_POOL_URIS / MAX_POOL_TOKENS
  • Erlang BEAM β†’ max_chunks
  • EVM bytecode β†’ EVM_INTERFACE_LIMITS.inputBytes (approach)
  • PCAP β†’ TSHARK_JSON_MAX_BYTES
  • GGUF β†’ MODEL_METADATA_NESTING
  • Electron ASAR β†’ DEFAULT_ASAR_MAX_HEADER_BYTES
  • Electron fuses β†’ ELECTRON_FUSE_LIMITS.maximumSentinelsPerSlice

Install (once published)

npm i -g [email protected]
# or from the GitHub Release tarball

Compatibility

  • Node.js ^22.23.4 (primary CI / engines) and 24.x (SQLite limits + empirical probes). Node 24.22.1 verified green this release.
  • Optional tooling for full empirical coverage: tshark, Wabt 1.0.42 (REA_WABT_BIN_DIRECTORY), sibling REA tree for verify:rea / redteam:rea.

Full changelog

See CHANGELOG.md.