# AEGIS — Threat Simulation Framework

- **Document ID:** ARCH-39
- **Phase:** E.2 — Verification Methodology
- **Status:** Draft for review (post four-reviewer discipline)
- **Version:** 0.1
- **Date:** 2026-07-10
- **Owner:** Chief Security Architect

---

## 1. Purpose

Define reproducible synthetic threat scenarios that verify AEGIS's detection, correlation, AI, response, and audit behavior against a curated adversary set. Implements the Living Threat Model (mandate #5) as a running system — threats are not a document, they are executable scenarios.

## 2. Model

- **Scenario Library.** Curated, versioned catalog of scenarios, each a structured YAML with: `id`, `title`, `MITRE_technique`, `description`, `preconditions`, `steps`, `expected_signals`, `expected_findings`, `expected_recommendations`, `scoring_criteria`.
- **Runner.** A dedicated `threatsim` runner spins up a test tenant + synthetic sources + expected outcomes; drives the scenario; captures outputs; compares to expected.
- **Scoring.** Detection latency, correct-classification rate, false-positive avoidance, evidence completeness, AI recommendation quality, audit-chain integrity.

## 3. Scenario Categories

Aligned to ARCH-03 adversary profiles:

- **Prompt-injection scenarios** (P-1). Attacker-shaped log lines; verify Safety Layer detection + evidence citation refusal.
- **Insider abuse** (P-4, P-5). Simulated capability over-use; verify anomaly detection + audit clarity.
- **External-network** (P-3). Simulated intrusion patterns; verify correlation rules + timeline construction.
- **Supply-chain** (P-6, P-7). Simulated compromised dependency behavior at controlled test-tenant; verify RIV alarms + trust decay.
- **Ransomware** (P-9). Simulated backup-target tampering; verify Recovery Domain isolation + air-gap copy integrity.
- **APT chain** (P-8). Multi-stage kill-chain; verify longitudinal correlation.
- **AI-specific** (P-1, P-10). Prompt-injection + model-backdoor scenarios; verify multi-model consensus + hallucination detection.

## 4. Reproducibility

- Every scenario is deterministic given signed inputs (`seed` + `input_corpus_hash`).
- Runs produce a **signed scenario report**: expected vs actual + evidence links.
- Reports feed the Verification Matrix (ARCH-37) as evidence.

## 5. Cadence

- **Per-release regression** — subset of scenarios (canary set) runs in CI.
- **Weekly** — full library on a test tenant.
- **Monthly** — new scenarios triaged and added.
- **Incident-driven** — post-mortem findings become new scenarios (mandate #5 Living Threat Model).

## 6. Ownership

- **Scenario Library owner:** SR pool with Threat-modeling contribution.
- **Runner owner:** platform testing team.
- **Reviewer:** Adversarial architect (per doc discipline) reviews new scenarios for coverage.

## 7. Independent Architecture Review

- **F-1.** *Scenario coverage may lag emerging threats.* Monthly triage + incident-driven addition + external red-team.
- **F-2.** *False confidence from passing scenarios.* Scenarios are one instrument among many; red team + audit calibrate.

## 8. Adversarial Architect Review

- **A-1.** *Attacker gams by memorizing scenarios.* Scenario library includes randomized variants; parameterization; new scenarios added continuously.
- **A-2.** *Insider disables a scenario.* Library versioned + signed; disables produce audit event + review.
- **A-3.** *Reproducibility limits realism.* Balanced by red-team engagements + chaos + purple team.

## 9. Operational Reliability Review

- **O-1.** Runner sustainable; per-tenant test infra hardened.
- **O-2.** Scenario debugging supported: report inspects step-by-step.
- **O-3.** 15-year sustainability: scenario schema stable; library grows.

## 10. Self-Critique

- **S-1.** *Scenarios written by us may miss what we haven't imagined.* Compensated by red team + external contribution + threat-intel-driven additions.
- **S-2.** *Deterministic scenarios could ossify detection to their exact shape.* Parameterization + fuzz-driven variants counter.

## 11. First-Target Analysis and Redesign

**Target:** the scenario library itself. If an attacker can quietly remove a scenario or dilute its expected outcomes, that class of behavior stops being detected. **Response:** library is signed configuration (ARCH-22 C-C class); ceremony changes; drift monitor; runs of removed scenarios still emit "expected but missing" reports.

## 12. Decisions

### D-39-1. Signed, reproducible scenario library + dedicated runner

- **Reason.** Living Threat Model as executable system.

## 13. Change Log

- **0.1 (2026-07-10)** — Initial draft.