Interactive tool
Injection Threat Modeler
Assemble your LLM application from building blocks. The modeler shows which of the five prompt-injection taxonomy classes become reachable, plus insecure output handling as a downstream consequence: which trust boundary each one breaks, the attacker access required, a sanitized proof-of-concept pattern (described, never a live payload), and the defenses that mitigate it, each annotated with its known failure mode. Export the whole threat model as Markdown.
Each of the five taxonomy classes becomes reachable only when your selected blocks expose its channel. Insecure output handling is listed alongside them as a downstream consequence, not a sixth class: it is what turns an injection that already landed into script or command execution in whatever renders the output. We describe sanitized PoC patterns only, never live payloads, in line with this site's stance. 5 taxonomy classes plus 1 downstream consequence · reviewed 2026-10.
The five-class taxonomy, plus output handling
| Class | Becomes reachable with | In one line |
|---|---|---|
| Direct prompt injection | user-chat, system-prompt | The user themselves supplies adversarial instructions. |
| Indirect prompt injection | rag-corpus, web-fetch, doc-ingestion | Instructions arrive through content the model consumes, not from the user. |
| Multimodal injection | doc-ingestion, web-fetch | The instruction is carried in a non-text modality (image/audio/file). |
| Agentic / tool-use injection | function-calling, multi-agent | Injection drives tool calls or hijacks downstream agents — leading to real-world actions. |
| Multi-turn / persistent injection | memory | The payload is split across turns, or planted in stored memory that the model re-reads later. |
| Insecure output handling downstream consequence | html-render | Not an injection channel: the downstream consequence when an injection's payload executes in whatever consumes the model's output. |