Keeping Secrets Out of Prompts and Logs

End-of-lesson check

15 questions · take it digitally for instant feedback at tendril.neural-forge.io/learn/quiz/end-tools-secret-handling-r12a1-creators

1. Which statement accurately describes a fundamental limitation of AI systems regarding secrets?

   1. AI models can reliably detect and refuse to repeat secrets they encounter in user input
   2. Once an AI model sees a secret, it will always refuse to output that secret in future responses
   3. AI models have built-in encryption that prevents logged conversations from being read by administrators
   4. AI systems cannot be trusted to keep secrets they have seen because prompts and outputs are logged and cached

2. A student accidentally pastes an API key into a prompt while testing a script. What does the lesson recommend as the correct response?

   1. Wait to see if anyone reports unauthorized access before taking action
   2. Assume the key is compromised and rotate it immediately, then investigate what systems logged the prompt
   3. Continue using the key since the AI model will not deliberately share it with anyone
   4. Only worry if the model explicitly outputs the key back in its response

3. What is the primary purpose of running user input through a 'redactor' before sending it to an AI model?

   1. To improve the model's response accuracy by removing confusing or irrelevant words
   2. To automatically generate better prompts by summarizing the user's intent
   3. To remove sensitive information like PII before it enters prompt logs and traces
   4. To format the input so it matches the model's preferred token patterns

4. A developer reasons that since their company owns both the AI model and the infrastructure, they can safely include customer Social Security numbers in prompts. Why is this reasoning flawed?

   1. Even in owned systems, prompts are logged, cached, and may be accessed by support staff or compromised in breaches— secrets should never enter prompts
   2. The model's training data will automatically encrypt SSNs for security
   3. Social Security numbers are not considered sensitive information under data protection laws
   4. The AI model will refuse to process prompts containing SSNs due to built-in safety filters

5. Which practice represents the MOST secure approach to handling user-provided credentials in an AI-powered application?

   1. Pass credentials through the prompt so the model can learn to recognize valid ones
   2. Keep credentials in your backend and reference them only by internal identifiers in prompts
   3. Store credentials in the system prompt for persistent access across sessions
   4. Let the user paste credentials directly into the chat so the model can validate them in real-time

6. The lesson mentions that secrets can 'sneak in' even with rules in place. What does this imply about security practices?

   1. AI models will automatically enforce rules better over time without human intervention
   2. Rules are useless and should not be implemented since secrets will leak anyway
   3. Once rules are set, no further action is needed to maintain security
   4. Rules combined with periodic scanning of logs for leaked data provide defense in depth

7. What makes 'deterministic scrubbing' more reliable than relying on the AI model itself to handle sensitive data?

   1. Deterministic scrubbing uses the same rules every time, while AI models can make inconsistent decisions based on context
   2. AI models are more thorough because they understand the meaning of text
   3. AI models automatically encrypt sensitive data before processing it
   4. Deterministic scrubbing is faster and cheaper than using an AI model

8. A developer integrates a third-party AI API and wants to minimize risk if that API experiences a data breach. Which strategy BEST addresses this risk?

   1. Avoid sending any user-provided data to the external API; use opaque identifiers instead
   2. Choose a different AI provider that claims to never log prompts
   3. Only send prompts during business hours when monitoring staff are available
   4. Add a disclaimer to users that their data may be logged

9. The lesson warns against pasting 'internal URLs' into prompts. Why are internal URLs considered sensitive in this context?

   1. Internal URLs contain encryption keys embedded in the domain name
   2. Internal URLs might reveal information about internal systems, naming conventions, or network structure that could aid attackers
   3. AI models cannot process URLs and will return errors
   4. Internal URLs are blocked by firewalls and cause errors if sent to AI APIs

10. What does it mean to treat prompts and traces as places where 'secrets leak by default'?

    1. Prompts are encrypted by default so secrets cannot be read
    2. The AI model automatically redacts secrets before they enter logs
    3. Assume any secret in a prompt will eventually be exposed through logs, caches, or shared systems
    4. Prompts should only be used for non-sensitive queries to prevent leaks

11. A student argues that they should never use AI APIs at all because all prompts are logged anyway. What's the flaw in this reasoning?

    1. Not all data is equally sensitive; the risk depends on what information enters the prompt
    2. Only government systems log prompts; commercial APIs do not
    3. AI APIs are never logged, so the student is wrong
    4. The student is correct— AI APIs should never be used for any purpose

12. Why might an organization want to periodically scan their prompt log stores for PII patterns, even if they have redaction in place?

    1. AI models cannot function without periodic scanning of their inputs
    2. Scanning is only useful if the organization plans to sell the log data
    3. Scanning helps identify cases where redaction failed or where new types of sensitive data appeared
    4. Redaction is perfect and never fails, so scanning is purely for compliance theater

13. A developer builds an internal tool where employees ask questions about company documents. The tool references documents by pasting full URLs into prompts. What is the BEST improvement to reduce security risk?

    1. Replace document URLs with internal document IDs that the backend resolves to actual content
    2. Use a different AI model from a competing provider
    3. Require employees to log in before asking questions
    4. Restrict the tool to only accepting questions during work hours

14. Which scenario represents the HIGHEST risk of secret exposure through AI prompt logging?

    1. Asking an AI to summarize a publicly available news article
    2. Asking an AI to write a creative story about dragons
    3. Including an API key in a prompt to test a connection
    4. Pasting a customer support ticket number into a prompt to retrieve customer details

15. A company implements a policy that all prompts must use internal customer IDs instead of names or account numbers. This is an example of which security concept from the lesson?

    1. Input validation
    2. Deterministic scrubbing
    3. Log rotation
    4. Opaque ID referencing