This repository contains examples of adversarial prompts designed to test AI systems for instruction-following, reasoning, consistency, and response quality.
The purpose of these prompts is to identify weak points in model behavior by introducing layered constraints, ambiguity, edge cases, and format-specific requirements.
These prompt examples are designed to test whether an AI can:
- Follow multi-step instructions exactly
- Maintain consistency across constraints
- Handle edge cases without hallucinating
- Distinguish relevant from irrelevant information
- Produce responses in the required format
Prompt:
Answer this question in exactly six words:
What are the benefits of exercise?
Do not use any word containing the letter "e".
Why this is adversarial:
The prompt combines a content requirement, a word-count requirement, and a character-level restriction. This tests whether the model can follow multiple constraints at once.
What this tests:
- Instruction-following
- Constraint handling
- Output precision
Prompt:
List four fruits in alphabetical order.
Return the answer as a numbered list.
Do not use the letters A, E, or O anywhere in your response.
Why this is adversarial:
The prompt requires the model to satisfy both content and formatting rules while also obeying a letter restriction that may make the task impossible or expose careless generation.
What this tests:
- Formatting compliance
- Constraint conflict detection
- Attention to detail
Prompt:
A red car and a blue bicycle are parked outside.
The sky is cloudy.
What color is the bicycle?
Answer using one word only.
Why this is adversarial:
This prompt checks whether the model can ignore extra details and extract only the relevant information.
What this tests:
- Relevance filtering
- Reading accuracy
- Concise response generation
Prompt:
Write the word "triangle" backwards.
Do not explain your answer.
Return the answer in uppercase letters only.
Why this is adversarial:
This tests whether the model can transform text correctly while still following formatting requirements.
What this tests:
- Text transformation
- Instruction-following
- Output formatting
Prompt:
Summarize the article in two sentences.
Do not add information that is not in the article.
[No article provided.]
Why this is adversarial:
This prompt tests whether the model notices missing context instead of inventing content.
What this tests:
- Hallucination resistance
- Missing-context detection
- Response safety
When reviewing responses to adversarial prompts, common evaluation areas include:
- Accuracy
- Instruction-following
- Consistency
- Hallucination avoidance
- Formatting compliance
- Reasoning quality
These examples reflect real-world AI evaluation work involving prompt testing, quality review, and identifying model weaknesses under difficult or conflicting instructions.