threat-model

name	threat-model
description	AI-specific threat model covering prompt injection, data poisoning, PII leakage, excessive agency, and supply chain risks

Skill: /threat-model — AI System Threat Modeling

Trigger

User runs /threat-model followed by a component or system description.

Behavior

Run an AI-specific threat model. Cover both traditional security concerns AND AI-specific attack surfaces that standard threat models miss.

AI-Specific Threat Categories to Always Cover

1. Prompt injection — direct and indirect
2. Data poisoning — training and retrieval-time
3. Model extraction / IP theft
4. PII leakage — via prompt, context, or model memorization
5. Jailbreaking / policy bypass
6. Hallucination exploitation — adversarial queries designed to trigger false outputs
7. Supply chain — model providers, embedding APIs, vector DBs
8. Excessive agency — agentic systems taking unintended real-world actions

Output Format

Threat Model: [Component/System Name]

Scope: [what's in / out of scope] Trust boundary: [where does untrusted input enter?]

---

Threat Surface Map

#	Threat	Category	Likelihood	Impact	Mitigation
1			H/M/L	H/M/L

Top 3 Priorities

What to fix first and why.

Monitoring Recommendations

What to log, alert on, and review to detect these threats in production.