ABS Core v4.3.3

Changelog

Release history for ABS Core — Runtime Governance for AI Agents.

Changelog

v4.5.0 — Intelligence Sovereignty & Privacy by Design

Release date: 2026-05-17 Architecture Score: 4.6/10 → 10/10 (independent verification) Automated tests: 0 → 55 passing TypeScript packages: 6 → 22 Python modules: 3 → 13 Documentation pages: 160 → 174 published


Security Fixes — P0 (Critical)

#FileBeforeAfter
1CORTEX/ledger.pyEd25519 key stored in plaintext PEM (NoEncryption())BestAvailableEncryption(password) via CORTEX_SIGNING_KEY_PASSWORD + hard fail in production + chmod 0600
2CORTEX/octagon_server.pyUnix socket with no authentication — any local process could inject into CognitionServiceTokenAuthInterceptor gRPC + constant-time comparison + chmod 0600
3CORTEX/mcp_server.pyZero authentication on all 6 MCP tools_require_auth() on every tool + _api_key in all schemas + hard fail in production
4CORTEX/app/service.pyCHI vaccination with silent fallback — gateway failure allowed unsanitized text into memoryCHIVaccinationError + CHI_VACCINATION_FAILURE ledger entry + total block (SI-004)
5CORTEX/sensory.pyInbox data/inbox/ ingested files without CHICHI gate before memorize() — failure = SENSORY_REJECTED in ledger
6OID/src/key-provider.tsstoreKeys() wrote PEM without encryptionaes-256-cbc + ABS_KEY_PASSWORD + hard fail in ABS_ENV=production

Security Fixes — P1 (High)

  • Mythos Gate 07 real logicAnalyzeIntent() was returning hardcoded ALLOW → replaced with real Mythos scorer (42 patterns)
  • PII vaccination real patternsVaccinate() used text.replace("CPF","[REDACTED]") → replaced with 5 compiled regex patterns
  • QUORUM hard fail — missing token was a silent pass → HTTP 503 in production + timingSafeEqual()
  • LEDGER constant-time auth!== comparison (timing attack) → timingSafeEqual()
  • Method collision resolvedmemorize(text, user_id) was overriding MemorizeMixin.memorize() (7-step pipeline inaccessible) → renamed memorize_simple()
  • retrieve() return type fixed — returned list[dict] but callers expected dict.contextretrieve_context() added

Security Fixes — P2 (Medium)

  • CHI vaccines — CPF + email only → 7 patterns: CPF, CNPJ, email, BR phone, card, RG, IPv4
  • behavioral_intuition() — was an empty stub → implemented with UEBA + LLM
  • start_mcp.sh — hardcoded /Volumes/LexarAPFS/CORTEX path → relative path

New TypeScript Packages (15 new)

PackageDescription
@abs-core/intelligenceUnified LLM+vector abstraction. Adapters: OpenAI/Azure, Anthropic (prompt caching), Ollama (local/air-gap), Cloudflare Vectorize, pgvector, in-memory. Single interface, swap via env var.
@abs-core/feature-flagsKV-backed feature flags with per-tenant/per-agent overrides and percentage rollout. Zero external dependency.
@abs-core/ipc-hmacHMAC-SHA256 inter-pillar auth with 60s replay protection + constant-time comparison (SI-003). Wired into LEDGER, QUORUM, CHI.
@abs-core/service-bindingsCloudflare Service Bindings — platform-native inter-Worker auth, zero shared secret, ~0.1ms latency.
@abs-core/policy-feedbackAutonomous policy learning: dissonances → LLM synthesis → human review queue → Mythos DB update.
@abs-core/tsa-internalSelf-hosted RFC 3161 TSA — eliminates FreeTSA.org dependency for air-gapped deployments. Integrated as primary fallback in LEDGER.
@abs-core/fipsFIPS 140-3 enforcement: CNSA 2.0 algorithm whitelist, OpenSSL FIPS mode check at startup, AES-256-GCM + HKDF-SHA384.
@abs-core/dod-pkiDoD PKI X.509 mTLS + CAC/PIV: certificate validation against DoD CA 3/4, CNSA 2.0 TLS ciphers, Hono middleware.
@abs-core/threat-intelSTIX/TAXII client: MITRE ATT&CK + CISA AIS + CISA KEV. Filters Chinese APT TTPs (Volt Typhoon, APT40, APT41). Auto-translates to Mythos candidates. Hourly sync on CLI startup.
@abs-core/correlationCross-deployment campaign detection. Rules: Volt Typhoon (30d), Salt Typhoon (7d), APT40 (14d), Coordinated Exfil (48h). STIX Campaign export. Wired into every LEDGER /record.
@abs-core/honeypot3 pre-built AI agent honeypots (finance-reporting-agent, hr-personnel-query-agent, infrastructure-inventory-agent). Watermarked synthetic data. Zero false positives. STIX Indicator per hit.
@abs-core/siemFormats: CEF (ArcSight), LEEF (QRadar), Syslog RFC 5424, Datadog JSON. Direct APIs: Darktrace, CrowdStrike Falcon, Splunk HEC. Wired into every gate decision.
@abs-core/pqcML-DSA-65 (CRYSTALS-Dilithium, NIST FIPS 204). Hybrid mode: Ed25519 + ML-DSA. Migration path by ABS_IMPACT_LEVEL. Addresses CNSA 2.0 NSS mandate for 2030.
@abs-core/model-integrityBehavioral fingerprinting (SHA-256 of 5 deterministic probe responses). Detects model tampering and substitution. ZK Commitment Scheme for classified audit.
@abs-core/edgeStandalone WASM governance for drones/IoT/submarines. Zero connectivity. Ed25519-signed policy bundles. Circular buffer ledger (1,000 decisions). Offline sync to central LEDGER.

New Python Modules (6 new)

ModuleDescription
cortex_engine/fips.pyFIPS 140-3 enforcement. CNSA 2.0 whitelist. validate_fips_at_startup() called on MCP server startup.
cortex_engine/ueba.pyLongitudinal behavioral analytics (30+ day baselines). Z-score anomaly scoring. Slow-burn detection (Volt Typhoon signature). Wired into behavioral_intuition() and octagon_server.py.
cortex_engine/privacy.pyDifferential Privacy ε=1.0 (Laplace noise on similarity scores). k-Anonymity validation. Cryptographic erasure with ZK proof (LGPD Art.18 / GDPR Art.17). Wired into retrieve_context().
cortex_engine/mythos_scorer.pyPython bridge to mythos-patterns.yaml (42 patterns). Phase 1: kill-switch < 500ms. Phase 2: scoring. Wired into octagon_server.py and intuition.py.
cortex_engine/intelligence_adapter.py@abs-core/intelligence pattern in Python. Adapters: OpenAI, Anthropic (prompt caching), Ollama, Mock. Replaces CortexLLMClient.
cortex_engine/ipc_hmac.pyIPC-HMAC Python library: sign/verify, replay protection, FastAPI dependency, IPCAuthError. 13 automated tests.

Package Wiring (6 packages connected to execution path)

ConnectionBeforeAfter
SIEM → LEDGERNever calledsiemRouter.route() after every /record
Correlation → LEDGERNever calledcorrelationEngine.ingest() after every /record
ThreatIntel → CLINever calledstartThreatIntelSync() on gateway startup
TSA internal → LEDGERFreeTSA.org as primaryHybridTSAService(internalTSA, externalFallback)
UEBA → octagon_serverNever called_feed_ueba() after every AnalyzeIntent()
@abs-core/intelligence → CORTEXHardcoded CortexLLMClient (OpenAI only)create_intelligence_provider() via LLM_PROVIDER env var
Privacy DP → retrieveNever calledadd_dp_noise_to_scores() in retrieve_context()
Policy Feedback → IntuitionEngineNever calledHigh-confidence dissonances sent to feedback collector

Privacy & Compliance (LGPD / GDPR / EU AI Act 2026)

  • Differential Privacy ε=1.0 — Laplace noise on all similarity scores in retrieve_context(). Prevents membership inference attacks.
  • k-Anonymity — Batch retrieval suppresses results when count < threshold (default: k=1 for retrieval, k=5 for exports).
  • Cryptographic erasure — Right-to-be-forgotten: content replaced with SHA-256 proof of existence. Preserves ledger chain integrity. LGPD Art.18 / GDPR Art.17.
  • Per-type memory TTL — PII: 90 days. Events: 1 year. Knowledge: permanent. Skills/tools: permanent.
  • ZK Audit Proofs — Prove governance happened without revealing TS/SCI or LGPD-protected payload. Pedersen commitment scheme.

DoD / Pentagon Hardening

  • FIPS 140-3 enforcement (code) — startup validation, prohibited algorithm blocking, CNSA 2.0 cipher suite enforcement
  • TPM 2.0 key provider — PCR sealing (PCRs 0,1,7,11), boot state attestation, hardware root of trust
  • PKCS#11 HSM abstraction — Thales Luna, Entrust nShield, YubiHSM support
  • DoD PKI X.509 mTLS — certificate chain validation, CAC/PIV, CNSA 2.0 TLS 1.3 ciphers
  • SELinux policiesabs_cortex.te + abs_ledger.te with neverallow rules
  • STIG Ansible playbook — FIPS, SELinux enforcing, kernel hardening, auditd rules, systemd service hardening
  • MLS Bell-LaPadulamls_bell_lapadula.rego: Simple Security + Star Property + CUI flow control + compartment handling (14 tests)
  • CUI classification — LEDGER labels per NIST SP 800-171, portion marks, cuiAuditMiddleware()
  • System Security Plan — Draft IL2 per NIST SP 800-53 Rev 5. POA&M with 8 items. Continuous monitoring plan.
  • IL4 Air-Gap Deployment Guide — 8-step guide: FIPS, TPM, DoD PKI, local infra, env config, STIG, deploy, validation

DevSecOps

  • Root git repo + .gitignore on all 9 pillars
  • @changesets/cli for coordinated multi-package releases
  • devcontainer.json + Docker Compose (pgvector:pg16, Redis) + automated setup.sh
  • Makefile — 25+ targets: make test, make migrate, make deploy-all, make sbom
  • uv Python package manager (10x faster than pip)
  • Drizzle ORM migrations for LEDGER (schema.ts + migration 0002: tenant_id, outbox, usage counters)
  • Alembic migrations for CORTEX PostgreSQL (pgvector schema)
  • SBOM blocking CI — CycloneDX 1.6 + Cosign keyless attestation on every PR merge
  • buf breaking — Protobuf contract breaking change detection blocks merge
  • E2E pipeline CI — all 11 gates tested on every merge to main
  • Pact contract tests — consumer-driven HTTP contract for gateway→LEDGER
  • OTLP batch export — Grafana Tempo, Honeycomb, Datadog. W3C Trace Context propagation. honoTracingMiddleware.
  • Multi-language SDK generation — buf.gen.yaml generates TypeScript, Python, Go from Protobuf
  • Renovate — automated dependency updates with security patch auto-merge
  • OPA policy unit tests + default-deny CI validation (root_policy_test.rego)
  • Per-tenant KV rate limiting (Cloudflare KV, standard HTTP headers)
  • Stripe metered billing (OCS/workers/billing.ts — USD 0.001/decision, 10k free tier/month)
  • Outbox pattern — LEDGER/src/event-publisher.ts: fan-out to Cloudflare Queues + webhook

Horizon 2030

  • Post-Quantum Cryptography — ML-DSA-65 (CRYSTALS-Dilithium, NIST FIPS 204). Hybrid mode (Ed25519 + ML-DSA). Migration modes by ABS_IMPACT_LEVEL. Addresses CNSA 2.0 NSS mandate.
  • Model Integrity — Behavioral fingerprinting detects backdoors and model substitution without access to weights.
  • ZK Audit Proofs — Prove governance over classified content without revealing it. TS/SCI + LGPD-compliant.
  • Edge Governance — Standalone WASM for drones, IoT, submarines. No server connectivity required.

Tests

55 automated tests (55 passing, 1 skipped):

SuiteTestsCovers
test_ledger.py5Hash chain, key encryption, production enforcement
test_intuition.py7Dissonance detection, behavioral fingerprint, malformed JSON
test_metabolism.py5REM consolidation, mirror write, ledger entry
test_mythos_scorer.py13Kill-switch, scoring, parameter patterns, decision matrix
test_ipc_hmac.py13Signing, verification, replay protection, tamper detection
test_octagon.py12PII vaccination, risk scoring, constant-time auth

Trial Sandbox (v4.5.0)

Expanded from 10 to 15 tests. New tests added:

  • Test 10 — Post-Quantum Cryptography ML-DSA-65 (NIST FIPS 204)
  • Test 11 — UEBA longitudinal analytics / Volt Typhoon slow-burn detection
  • Test 12 — Model integrity behavioral fingerprinting
  • Test 13 — Zero-Knowledge audit proof (governance without payload reveal)
  • Test 14 — Edge governance WASM (standalone, no connectivity)

Package SHA-256: c4c4e6355c891f8e0104426d99a2c75b21850c4e6cbb359e399132b9b7c525c1

Documentation

13 new or fully rewritten pages in docs.abscore.app:

changelog · welcome (v4.5.0 section) · security/ipc-hmac · security/hardening-v433 · technical/cortex-worker · technical/mythos-patterns · technical/pqc · technical/ueba · technical/threat-intel · technical/correlation · technical/honeypot · technical/siem-integration · technical/model-integrity · technical/edge-governance · guides/il4-airgap-deployment · investors/roadmap-2030


v4.3.3 — Sovereign Defense & Zero Trust

Release date: 2026-04-17

  • 11-gate governance pipeline: OID_VERIFY → CONTEXT_FENCE → OPA_FIREWALL → CANARY → TOOL_VALIDATE → TEMPORAL_FENCE → ADVERSARIAL → HUMAN_QUORUM → LEDGER_COMMIT → SECURE_SANDBOX → FINAL_AUDIT
  • RFC 3161 timestamping (FreeTSA.org + DigiCert fallback)
  • WASM Behavioral Scorer — Rust evaluate_mythos_patterns() + TypeScript AdversarialDetector bridge
  • M-of-N Human Quorum Gate with Redis HA (tiered: LOW 1/2, MID 2/3, HIGH 3/4)
  • MCP Gateway proxy (@oconnector/mcp-gateway) — intercepts all tool calls
  • CORTEX Memory Engine (metabolism, rehydration, dissonance detection)
  • Cloudflare D1 sovereign memory backend
  • Ed25519 hash-chain ledger with SHA-256
  • Protobuf contracts as single source of truth (buf.build)

v4.2.0 — Octagon Architecture

  • Industrial Octagon: OID, AICCP, QUORUM, CHI, LEDGER, ARCHAEO, CORTEX, OCS
  • Protobuf contracts as single source of truth
  • OPA/Rego policy engine compiled to WASM
  • Air-gapped deployment mode
  • pnpm monorepo workspace

v4.0.0 — Foundation

  • ABS Core runtime governance engine
  • Ed25519 identity verification (Gate 01)
  • Deterministic WASM policy evaluation (Gate 03)
  • SHA-256 hash-chain ledger
  • MCP protocol support
  • On-premise / air-gapped deployment bias

On this page