Review my customer-facing chatbot for security risks
Threat review for an AI support bot covering prompt injection, data leaks, abuse, and unsafe actions, with prioritized fixes and tests.
The prompt
Review the security of my customer-facing AI chatbot.
What the bot does: [e.g. "answers order questions and can issue refunds under $50"]
Channels: [website widget / WhatsApp / app / email]
Model and setup: [hosted model, retrieval from a knowledge base, tools or APIs it can call]
Data it can see: [order history, account details, internal docs, payment data]
Actions it can take: [read-only / update records / send messages / refunds]
Current safeguards: [authentication, rate limits, filters, human review, logging]
Assess these threat areas for my setup:
1. Prompt injection - direct (user messages) and indirect (content in documents, web pages, or emails the bot reads).
2. Data exposure - leaking other customers' data, system instructions, internal docs, or secrets.
3. Unsafe actions - tool calls a user could trigger without proper identity checks or limits.
4. Abuse and cost - spam, scraping, denial of wallet, harassment, off-topic use.
5. Harmful or off-brand output - false promises, policy commitments, offensive content.
6. Logging and privacy - what is stored, who can see it, retention.
Return:
- A risk table: Threat | How it could happen here | Likelihood | Impact | Fix.
- The top 5 fixes in priority order (e.g. least-privilege tool access, verifying identity before account actions, confirmation steps, output filtering, monitoring).
- 10 test prompts or scenarios I can run safely against my own bot to check each risk, with the expected safe behavior.
- What to monitor after launch and when to bring in a security professional.
Only suggest testing systems I own or am authorized to test.
Highlighted [brackets] are blanks to replace with your own details.
Fill in the blanks and use this promptMore Software Development prompts
What to fill in
- website widget / WhatsApp / app / email
- order history, account details, internal docs, payment data
- read-only / update records / send messages / refunds
- authentication, rate limits, filters, human review, logging
How this prompt is built
This prompt follows the Few-Shot structure (Few-shot prompting). A few-shot prompt includes one or more worked examples of the input and the output you want, so the model can copy the pattern instead of guessing it.
Related prompts
- Analyze logs to find what went wrongReads logs to identify errors, patterns, and the likely root cause.
- Get a beginner-friendly code review of my first projectA ranked set of the five improvements that matter most, explained simply, plus what you did well and what to learn next.
- Turn a threat report into Sigma detection rulesExtracts behaviors from a threat report, maps them to MITRE ATT&CK, and writes tuned Sigma rules with test notes.
- Review every change since my last release before shippingA pre-release review of all changes since a tag: missing translations, simplifications, breaking changes, security, and a draft changelog.