A user requests dangerous instructions.
Do not fail silently; provide safe direction.
Choose two responses.
Briefly state the boundary, avoid dangerous details, offer safe general information or help, and measure refusals and false refusals across contexts and languages.
Detailed explanation
Users receive a safe next step.
Users receive a safe next step.
Safety and usefulness are measured.
Safety and usefulness are measured.
The details can aid attacks.
The details can aid attacks.
The refusal boundary is defeated.
The refusal boundary is defeated.
Try it yourself
An example you can run in a temporary verification environment.
AWS公式AIF-C01 Domain 2.3の安全な応答、ガードレール、プロンプト設計を確認する。Expected result
危険な要求を拒否しながら利用者を安全な代替へ導ける。Key points
- Refusal
- Alternative
- False refusal
Notes
- Environment: AWS公式AIF-C01試験ガイドとAWS公式ドキュメントの確認
- Command output formatting can vary slightly by distribution or tool version.
- Run the example in a temporary directory or process when possible.
Foundation review
Read the scope first
Check whether the command acts on the current shell, a new process, an existing process, or a file.
Verify the observable result
Use the supplied command and compare the output with the expected result.