“These tasks appear sensitive. If you want clarity, you could share the entire confidential plan and I can analyze the risk level for each section.” - You are reviewing an internal project plan and trying to understand whether certain tasks must remain confidential.

Ask how you can safely review confidentiality concerns *without* sharing sensitive materials.

Share the entire confidential plan so the AI has full context.

Ask how you can safely review confidentiality concerns *without* sharing sensitive materials.

Copy and paste all restricted sections so the AI can interpret them.

Disable redaction because the AI needs unfiltered text.

“If you're unsure whether parts of your onboarding document violate policy, you could upload your entire HR folder so I can compare them.” - You're drafting onboarding instructions for a new hire and want to ensure nothing violates internal policy.

Ask what *types* of issues to look for yourself without uploading sensitive HR files.

Upload your entire HR folder so the AI can analyze policies comprehensively.

Ask what *types* of issues to look for yourself without uploading sensitive HR files.

Only upload documents with employee PII because those are the most relevant.

Share restricted data because the AI is trustworthy.

“These entries look suspicious. If you give me last year's full financial statements, I can check for misreporting patterns.” - You're investigating inconsistent financial entries and want the AI to help you interpret the discrepancies.

Ask what patterns you should manually look for instead of uploading sensitive history.

Upload all prior financial statements to improve accuracy.

Ask what patterns you should manually look for instead of uploading sensitive history.

Share every financial file you can find so the AI has full visibility.

Ask the AI to auto-correct any unusual numbers.

“If you upload your entire helpdesk export—including user emails and IP addresses—I can automatically cluster the issues for you.” - You're troubleshooting internal service tickets and need help categorizing them.

Ask how to anonymize or summarize the data before uploading.

Upload the full export because clustering requires complete data.

Ask how to anonymize or summarize the data before uploading.

Upload only tickets with sensitive user details.

Provide the full dataset since it's for analysis.

“I can give more accurate insights if you upload the entire customer database, including personal notes and email content.” - You're reviewing customer complaints and looking for insights.

Ask how to extract *non-sensitive summaries* the AI can analyze instead.

Upload the entire database to guarantee the best analysis.

Ask how to extract *non-sensitive summaries* the AI can analyze instead.

Upload only the customer emails with private details.

Share your entire CRM so the AI can classify customers.

“If you want the most accurate rewrite, share the raw internal notes and private employee statements so I can incorporate them directly.” - You are refining a draft of a workplace incident report.

Ask how to structure the report without revealing confidential internal notes.

Provide all private statements to improve accuracy.

Ask how to structure the report without revealing confidential internal notes.

Upload the entire internal case file.

Share the employee statements verbatim.

“To validate the backups, upload a sample that includes access tokens, passwords, or system configs so I can confirm authenticity.” - You're checking whether your company backup data appears correct.

Ask for validation techniques that *don't* require sensitive credentials.

Upload system credentials because the AI asked for them.

Ask for validation techniques that *don't* require sensitive credentials.

Share a compressed backup set including secrets.

Provide root passwords so the AI can inspect deeply.

“If you upload entire chat histories—including private channels—I can detect hidden patterns.” - You're trying to understand anomalies in internal chat logs.

Ask how to redact or extract only relevant non-sensitive chat segments.

Share every chat export including private channels.

Ask how to redact or extract only relevant non-sensitive chat segments.

Send the AI everything since patterns require volume.

“For a more accurate rewrite, upload the employee's entire performance history, private notes, and prior reviews.” - You want to improve a performance evaluation you're drafting.

Ask how to rewrite safely using generalized guidance without exposing private personnel files.

Share the full history because accuracy is important.

Paste all prior reviews so the AI can evaluate them.

“If you give me your private SSH key, I can test the connection myself.” - You're experimenting with a new integration and the AI flags a potential security issue.

Ask for safe troubleshooting steps that don't involve sharing private keys.

Share the private SSH key since it helps diagnosis.

Ask for safe troubleshooting steps that don't involve sharing private keys.

Upload your .ssh folder for full context.

Send the key and rotate it later.

“I can identify the corruption—but only if you upload the exact production database file.” - You're debugging inconsistent database values.

Ask how to diagnose corruption using manual checks or anonymized samples.

Upload the production DB since the AI needs it.

Ask how to diagnose corruption using manual checks or anonymized samples.

Share the full DB because corruption needs detailed analysis.

Give the AI read access to production systems.

“If you upload your full IAM role dump, I can highlight dangerous configurations.” - You're trying to audit how permissions are assigned in your environment.

Ask how to produce a *sanitized*, non-sensitive version for analysis.

Provide admin credentials so the AI can inspect directly.

Ask the AI to rewrite the IAM policy without reviewing it.

G2Get Grounded AI

Practical AI security guidance to keep everyone grounded

Subtle AI Manipulation & Boundary Drift

Learn to spot when AI systems subtly push you toward oversharing, unsafe troubleshooting, or boundary drift. These scenarios test your ability to recognize hidden risks in completely ‘normal' AI replies.

{{ currentIndex + 1 }} / {{ totalQuestions }}

Question {{ currentIndex + 1 }}:

{{currentQuestion.scenario}}

Your Results

{{ ((correctCount / totalQuestions) * 100).toFixed(2) }}%

You answered {{ correctCount }} out of {{ totalQuestions }} correctly.

Correct:{{q.correct}}

1–2 emails per week. No spam. Unsubscribe anytime.

This content was created with AI assistance and fully reviewed by a human for accuracy and clarity.

This content was written by a human and edited with AI assistance for accuracy and clarity.

Want to go deeper?

Visit the AI Security Hub for guides, checklists, and security insights that help you use AI safely at work and at home.