Skip to content

threat-system-architecture-spec.md

Latest commit

 

History

History
89 lines (71 loc) · 4.96 KB

threat-system-architecture-spec.md

File metadata and controls

89 lines (71 loc) · 4.96 KB
schema_version 1.0
template system-architecture
date 2026-04-10
source_file compensating-controls.md
finding_count 22
image_generated false

1. Metadata

Field Value
Project Name Agentic AI Application
Scan Date 2026-04-10
Analysis Agents 8 (S, T, R, I, D, E, AG, LLM)
Total Findings 22
Risk Posture Residual risk — 0 Critical and 10 High findings across 7 components

2. Risk Distribution

Severity Count Percentage Color
Critical 0 0.0% #DC2626
High 10 45.0% #EA580C
Medium 12 55.0% #CA8A04
Low 0 0.0% #2563EB
Total 22 100% --

Chart Format: Suitable for donut chart (proportional segments) or horizontal bar chart (comparative lengths).

3. Coverage Heat Map

Component Critical High Medium Low Total
LLM Agent Orchestrator 0 4 6 0 10
MCP Tool Server 0 2 2 0 4
Guardrails Service 0 2 0 0 2
Knowledge Base 0 0 2 0 2
User 0 1 1 0 2
Audit Logger 0 0 1 0 1
External API 0 1 0 0 1

4. Top Critical Findings

# Finding ID Component Threat Risk Level
1 LLM-1 LLM Agent Orchestrator Indirect prompt injection via documents retrieved from the … High
2 AG-4 MCP Tool Server Compromised or manipulated agent triggers excessive tool in… High
3 E-2 Guardrails Service Attacker bypasses guardrails validation through prompt obfu… High
4 E-1 LLM Agent Orchestrator Orchestrator escalates its own tool permissions beyond the … High
5 LLM-3 LLM Agent Orchestrator Attacker crafts prompts that cause the LLM to generate tool… High

5. Architecture Threat Overlay

Component Risk Weight Finding Count Annotation
LLM Agent Orchestrator High 10 Central coordination hub dispatching LLM inference, tool calls, and knowledge retrieval. Concentrates the majority of findings as the primary target for spoofing, tampering, information disclosure, privilege escalation, and AI-specific (LLM + agentic) threats.
MCP Tool Server Medium-High 4 Executes tool calls on behalf of the orchestrator. Exposed to prompt-driven parameter manipulation, unauthorized tool invocation, resource exhaustion, and scope-crossing agentic actions.
Guardrails Service Medium-High 2 Entry-point input filter screening user prompts. High-severity denial-of-service and privilege-escalation findings reflect its position as the first attacker-facing component in the application zone.
Knowledge Base Medium 2 Vector store providing RAG context. Data integrity (embedding poisoning) and confidentiality (embedding reversal) findings reflect its role as the indirect prompt injection attack surface.
User Medium-High 2 External entity submitting prompts. Spoofing (token replay) and repudiation (insufficient session attribution) findings reflect identity-boundary risks.
Audit Logger Medium 1 Centralized observability and forensics store (MAESTRO L5). Tampering findings reflect accountability risks if log integrity is not cryptographically chained.
External API Medium-High 1 Third-party service reached through the tool server. Spoofing findings reflect response-integrity risks at the trust-zone boundary.

6. Visual Design Directives

Format: Portrait orientation (2:3 aspect ratio), print-ready at 300 DPI, 2480 x 3508 px minimum.

Color System:

  • Critical: #DC2626 (deep red) — urgent action required
  • High: #EA580C (orange-red) — priority remediation
  • Medium: #CA8A04 (gold) — scheduled remediation
  • Low: #2563EB (blue) — backlog/monitor
  • Background: light gray #F5F5F5 with white panels #FFFFFF
  • Headers: dark navy #1E293B
  • Accent lines: medium gray #64748B

Layout:

  • Top banner (12% height): project name, scan date, total findings
  • System architecture diagram (60% height, center): components as nodes, data flows as edges, trust boundaries as zone backgrounds
  • Risk annotations per component overlay (finding counts in colored bubbles)
  • Bottom panel (28% height): top findings legend + risk distribution bar chart

Typography: Same as baseball-card. Component labels in 12pt bold; data flow labels in 9pt italic.

Visual Style: Technical diagram aesthetic. Use flowchart conventions (boxes for processes, cylinders for data stores, parallelograms for external entities). Trust zones shown as rounded-rectangle backgrounds with subtle fill colors.

Prompt Hint (Gemini): Generate a print-quality infographic using the layout above. Emphasize canonical MAESTRO layer naming — use ONLY the canonical CSA MAESTRO layer names: L5 Evaluation and Observability, L6 Security and Compliance, L7 Agent Ecosystem.