Zum Hauptinhalt springen

Worked Walkthrough: Threat-Model a Prompt Injection Path

Run a simple, repeatable prompt-injection threat-model on an AI workflow.

ai-security-safety6 min read

A research agent can search the web, read documents, and email a summary to a client account manager. Map untrusted input → model interpretation → available action → guardrail. Teams often stop at “the model might be manipulated” and never trace which real action could follow from that manipulation. 1. Mark entry points List every untrusted source the model can read: search results, webpages, PDFs, ticket notes, attachments. If the source can contain attacker text, it belongs in the threat model even if users think of it as “content,” not “input.” 2. Mark privileges Write down the actions still available…

Read the full lesson

Sign up free: one personalized lesson every day, matched to your role and goals.

← Back to library

Kontakt aufnehmen