365 Architect

Aegis 365 -- Testing & Quality Assurance

Test Level Coverage Tool Target
Unit L0--L7 components; crypto, encoding xUnit, Moq, FluentAssertions > 85% coverage
Integration L0--L7 pipeline; DB, cache, SLM TestContainers, In-Memory DB > 80% scenario coverage
Security Injection, re-ID risk, encryption OWASP ZAP, custom exploit suites 100% known attacks detected
Resilience Failover, latency, fail-closed Chaos Monkey, load tests, fault injection RTO/RPO met
Compliance Audit completeness, Break Glass, immutability Manual + automated verification Zero audit gaps
Performance L0--L7 < 2.5s (human), < 5s (agentic) JMeter, BenchmarkDotNet 450ms aspirational
Penetration Red-team, supply-chain External firm (Phase 5) No critical findings

Unit Testing Strategy

Every layer component is tested in isolation with mocked dependencies:

  • L0: Action taxonomy classification, HITL routing logic, session state transitions
  • L1: SLM confidence scoring (known-attack corpus must block), transparency model routing
  • L2: Regex + NER + SLM consensus paths, Private Semantic Map matching
  • L3: Token generation collision tests, SDG plausibility verification
  • L4: Re-hydration hierarchy (clearance checks, Hard Mask enforcement), mask-on-hover audit trigger
  • L5: Compliance registry lookup, sovereignty precedence ordering
  • L6: ZKP generation correctness, Break Glass dual-key enforcement, immutable chain integrity
  • L7: Semantic hash stability, vector distance thresholds, inference-attack suppression

Integration Testing Strategy

End-to-end pipeline tests run with TestContainers (Redis, PostgreSQL) and an in-memory SLM stub:

  1. Ingest a prompt with known PII
  2. Verify L0 action classification, L1 intent score, L2 sensitivity tile
  3. Verify L3 anonymization produces expected token pattern
  4. Verify L5 routing respects data classification
  5. Verify L4 re-hydration returns real value (with clearance) or mask (without)
  6. Verify L6 audit log contains immutable chain entry with ZKP witness
  7. Verify L7 cache stores anonymized response only, with correct TTL

Security Testing Strategy

Attack Vector Test Method Success Criteria
Prompt injection / jailbreak Known-attack corpus (100+ patterns) > 99% blocked by L1
Indirect prompt injection Poisoned external content ingestion L1 Instruction Integrity Check flags
Mosaic / drip-feed exfiltration Multi-session aggregation probe L0 temporal detection triggers freeze
Re-identification attempt Token combination analysis L3 collision-risk guard triggers generalisation
Inference de-masking LLM response probing L4 post-inference redaction intercepts
Admin plaintext bypass Single-credential decryption attempt MPC quorum required; alert generated
Memory-wipe evasion Controlled node crash simulation Synchronous replication preserves state

Share on LinkedIn