AGENT SKILL SPECIFICATION // .agents/skills/ai-systems-sentinel/
● Mandatory Stage 2 Pipeline Gate
AI Systems Sentinel is a distilled, custom-brewed autonomous agent super-skill engineered to defend production Applied AI applications. Embedded directly into developer agent pairing sessions, it automatically halts builds and executes a rigorous 4-layer security audit before any frontend, API, container, or AI gateway update is permitted to deploy.
4-Tier Defense Architecture & Boundary Matrix
AUTONOMOUS AGENT SUPER-SKILL
AI-SYSTEMS-SENTINEL
Stage 2 Pipeline Gate // Full-Stack & AI Defense
- Zero-Trust CSP
- Anti-Clickjacking (DENY)
- DOM-XSS TextContent Ban
- Subresource SRI Hashes
- OWASP API Top 10
- Strict Pydantic Typing
- Path Traversal Defense
- SSRF Boundary Mesh
- Non-root USER 1000
- Zero Host Mount Overrides
- 127.0.0.1 Docker Bindings
- Cgroup Quotas (CPU/RAM)
- Prompt Injection Shields
- Zero Client Key Exposure
- DaveRouter Proxy Gateway
- Token-Burn Circuit Breakers
LAYER // 01
● Perimeter & Edge
Edge Perimeter & Web Defense
Strict Zero-Trust security headers enforced at the Cloudflare edge and application shell. Complete elimination of DOM-based XSS via safe textContent primitives and Jinja2 mandatory auto-escaping.
CSP
Strict Self
Frames
DENY
Storage
UI State Only
Zero-Trust CSP
Anti-Clickjacking
DOM-XSS Prevention
SRI Integrity
LAYER // 02
● API & Contracts
API & Gateway Armor
Strict Pydantic v2 payload validation rejecting mass-assignment/auto-binding. Path traversal prevention via canonical realpath boundary verification. Loopback-only routing preventing SSRF exploits.
Traversal
CWE-22 Block
Validation
Pydantic v2
SSRF
Boundary Mesh
OWASP API T10
Path Canonicalization
Type Enforcement
SSRF Defense
LAYER // 03
● Dave Homelab
Sovereign Bare-Metal Hardening
Container runtime defense-in-depth: non-root UID 1000 execution, immutable registry images with zero host volume overrides, Linux cgroup memory/CPU throttling, and internal loopback bindings.
Privilege
UID 1000 (Non-Root)
Cgroups
512M / 0.75 CPU
Exposure
0 Public Ports
Cgroup Throttling
Least Privilege
Zero Open Ports
Docker Compose
LAYER // 04
● Neural Inference
Applied AI & LLM Guardrails
Zero client-side LLM key exposure. All AI actions route via sovereign loopbacks with delimiter-tagged prompt containment, max_tokens output caps, and token-burn circuit breakers.
Keys
0 Client-Exposed
Injection
Tag Delimited
Breaker
Denial-of-Wallet
OWASP LLM T10
DaveRouter Proxy
Prompt Firewalls
FinOps Breakers