Works with the AI tools you already use

    CClaude CodeCCursorCCodex CLIGGitHub CopilotGGemini CLI+17 more

    AI Agent Threat Model — Prompt Injection & Tool Misuse

    by monna

    1

    Map trust boundaries, trace injection-to-tool attack paths, and prioritize concrete controls before an AI agent ships.

    Secure checkout via Stripe

    0 installsSecurity scanned

    See it in action

    You say

    Threat-model our support agent. It reads customer emails and RAG results, can search the knowledge base, send emails, and issue refunds up to $500 without human approval. We have not documented vector-store write validation or customer-level retrieval scoping.

    Your agent does

    VERDICT: NOT-DEPLOYABLE

    CRITICAL — Combined tampering + elevation path Untrusted email content reaches the model context while the refund tool can act without human review.

    Evidence: Stated Confidence: 88% — based mainly on stated autonomy and tool authority Priority control: require human confirmation for refunds, isolate untrusted content from instructions, and cap tool authority.

    What you get

    Audit RAG pipelines for indirect prompt injection and memory poisoning.Analyze agentic systems for unauthorized tool execution and privilege escalation.Map AI-specific vulnerabilities to concrete security controls and mitigations.Generate structured DFDs for AI applications to identify hidden trust boundaries.

    About this skill

    What it does

    An AI agent can turn untrusted text into a real action. This skill reconstructs the system as a data-flow diagram, applies STRIDE to every relevant element, and treats prompt injection plus high-impact tool access as one compound path instead of two disconnected checklist items.

    What you get

    • A deployment verdict and AI data-flow inventory.
    • Trust boundaries and a STRIDE-per-element matrix.
    • Evidence-tagged findings with confidence scores.
    • A mitigation roadmap ordered by consequence.
    • A deterministic coverage check for every applicable STRIDE category and the combined injection-to-tool path.

    Worked path included

    Send no architecture and the skill runs a labeled demonstration on its built-in customer-support agent. The package also includes a complete worked example and a universal prompt edition for capable chat models that do not load skills directly.

    Known limitations

    This is a design-stage threat model, not a penetration test or safety certification. It works from the architecture and permissions you describe, does not inspect a live system, and does not generate adversarial instruction payloads or working exploits. Classic web/API logic and cloud IAM require their own reviews.

    How to install

    Drop the file into your AI Agent. Works with Claude, Cursor, ChatGPT, and 20+ more.

    Reviews

    No reviews yet

    Be one of the first to try it. Every listed skill passes our trust checks below.

    Security scanned

    Passed our 8-point scan before listing

    Fresh listing

    Recently published to Agensi

    30-day refund

    Not a fit? Get your money back

    Frequently Asked Questions

    Popular in AI Agents & LLM Ops