prompt-injection-auditlisted

# Prompt-Injection Audit Help a developer find and fix prompt-injection weaknesses in an AI system they own or are authorized to test. The output is a **Findings Report** modeled on a professional security assessment: each finding has a category, severity, evidence, and a concrete fix. ## Authorized-use boundary (read first) This skill is for **defensive testing of the user's own system**. It maps attack *surfaces* and provides *probe templates* that reveal whether a defense holds — the same way a penetration-testing framework works. It does **not**: - Produce turnkey jailbreaks tuned to defeat a specific named production model or a third party's system. - Help exfiltrate data, escalate privilege, or attack infrastructure the user doesn't control. If a request shifts from "test my system" to "help me break into someone else's," stop and say so plainly. Confirm scope up front: *"I'll treat this as authorized testing of your own system — is that right?"* When in doubt, keep probes illustrative (do they bypass the guardrail: yes/no) rather than weaponized. ## Workflow ### Step 1 — Map the attack surface Inventory where untrusted text can reach the model. Ask for (or read from provided files) whichever apply: - **System prompt** — the instructions the developer wrote. - **Tool / function definitions** — names, descriptions, parameter docs (a common tool-poisoning vector). - **Retrieval sources** — RAG documents, web fetches, database fields, anything the model reads at r