List the ways an AI agent could fail and how to prevent each
A failure-mode review for an AI agent: realistic scenarios, how you would notice, the damage, and the safeguard.
The prompt
Run a failure-mode review on an AI agent before we launch it.
What the agent does: [TASKS and actions]
Systems and data it can access: [DETAILS]
Inputs it receives: [e.g. "customer emails", "uploaded invoices", "web pages"]
Safeguards already planned: [APPROVALS, LIMITS, LOGGING]
Generate realistic failure scenarios across these types:
- Wrong output - made-up facts, wrong numbers, misread documents, outdated information.
- Wrong action - right idea, wrong record, wrong recipient, duplicate actions, actions in a loop.
- Manipulation - instructions hidden in emails, files, or web pages (prompt injection), social engineering, fake urgency.
- Data exposure - sharing confidential or personal data with the wrong person or tool.
- Edge cases - unusual formats, missing data, other languages, requests outside scope.
- Silent failure - the agent stops working or degrades and nobody notices.
- Human factors - over-trust, rubber-stamp approvals, unclear ownership.
- Dependency failure - a connected system is down or changes its format.
Return a table: Scenario | Example in my process | Likelihood (H/M/L) | Impact (H/M/L) | How we would detect it | Prevention or safeguard | Owner.
Then list:
1. The top 5 scenarios to address before launch.
2. Test cases to run for each of those 5.
3. Any scenario serious enough that the agent should not take that action at all.
Be specific to my process rather than listing generic AI risks.
Highlighted [brackets] are blanks to replace with your own details.
Fill in the blanks and use this promptMore Working with AI prompts
What to fill in
- TASKS and actions
- DETAILS
- e.g. "customer emails", "uploaded invoices", "web pages"
- APPROVALS, LIMITS, LOGGING
How this prompt is built
This prompt follows the RTF structure (Role, Task, Format). RTF tells the model who to be, what to do and what shape the answer should take. It is the quickest structure that still gives noticeably better results than a bare request.
Related prompts
- Reverse-engineer the prompt behind an outputWorks backwards from text, code, or an image to the prompt that likely produced it.
- Answer my question, then critique it and argue against itGets an answer, an honest self-critique, the strongest counterargument, and a revised answer you can trust more.
- Turn a one-off AI chat into a reusable prompt pack for my teamExtracts the prompts, fixes, and checks from an AI chat that worked into a step-by-step workflow and prompt pack colleagues can reuse.
- Identify the decisions in a task I should not hand to AIPinpoints the decisions, ambiguities, and exceptions in a task that a person must own, and why AI should not decide them.