Create advanced Works with: Any AI

Prompt Injection Risk Review

Table output Has example

Check how untrusted text can hijack your prompt.

The Open buttons send the prompt text as currently shown below to that site only when you click. Fill in the placeholders first.

Review my prompt-based workflow for prompt injection risks. The workflow (what the assistant does, what text it reads from users, documents, websites, or emails, and what actions or outputs it can produce): [WORKFLOW_DESCRIPTION]. The system prompt: [SYSTEM_PROMPT]. Tools or data the assistant can access: [TOOLS_AND_DATA].

Return:
1. A data-flow table: Source of text | Trusted? | Where it enters the prompt | What the assistant can do next (answer, call a tool, send a message, edit data).
2. Attack scenarios (described conceptually, not as working exploits): an email that tells the assistant to forward data; a document containing hidden instructions; a web page that asks to ignore previous rules; a customer message asking to reveal the system prompt. For each, the potential impact and likelihood.
3. Mitigation checklist: separate instructions from untrusted content with clear delimiters and state that content inside them is data, not instructions; give the assistant the least access needed; require human approval for sensitive actions; validate outputs and tool calls in code; do not place secrets in prompts; filter or limit what is returned; log activity.
4. Test cases with fictional malicious text to try in a safe environment.
5. A residual risk statement: delimiters and instructions reduce but do not eliminate the risk.

Add a decision rule for which actions must never be automated. Have a security professional review high-risk deployments.

Fill in before use

Optional: type your details and the prompt above updates. Nothing is saved or sent.

Replace each bracketed placeholder with your own details, then paste into Any AI or another assistant.

Data-flow table
Source | Trusted? | Enters at | Next action
Customer email | no | user message | draft reply
Order database | yes | tool result | read-only
Web page text | no | summarization | summary only
Scenario (conceptual): an email contains text telling the assistant to forward the order list. Impact: data exposure. Likelihood: medium.
Mitigations: delimit untrusted content and call it data; no send-email tool without approval; validate tool calls in code; no secrets in prompts; logging.
Residual risk: instructions reduce but do not eliminate injection; never automate sensitive actions. Security review for high-risk use.

• prompt-injection