Red Team an Assistant

1. For the Assistant you want to red team, click on the three dots next to it.

2. Click on Red Team to open the red teaming tab.

3. Here you will be able to see the list of red team attacks with statuses Pending, Running, and Completed

4. Let’s create a new Attack by clicking on Create Attack.

5. Type a Name for the attack that reflects the scenario or risk area you want to test.

6. In the red teaming tab, choose the Attack Type you want to run

7. From the list of available Probes, select the probe set that matches your chosen attack type or scenario.

Each probe, converter, and detector will include a description outlining its use case to help you select the appropriate option for your scenario. For a complete guide on the different types of probes, converters, and detectors, view Assistant Red Teaming

8. Select the Converter that defines how prompts and responses will be transformed or formatted before being sent to the model.

9. Choose the Scorer that will evaluate the assistant’s responses.

10. Review your configuration: Attack Type, Probe, Converter, and Scorer to ensure they match your testing goals.

11. Click on the start attack button to start the red ream attack.

12. A new job will appear in the jobs table with status Pending. After a short time, the status will update to Running as the attack is executed.

13. Click on the Refresh Table button to update the job list and check for progress.

14. Once the attack completes, the job status will change to Completed.

15. Click on the completed job to open the detailed results view.

16. Here you can inspect the prompts, model responses, and Scorer outputs for each red team example.

17. You can also download the complete html report to view the vulnerable prompts (if any).

Use these results to identify failures, unsafe behaviour, or other issues, and decide on follow-up fixes for your assistant.

Red Team an Assistant

1. For the Assistant you want to red team, click on the three dots next to it.#

2. Click on Red Team to open the red teaming tab.#

3. Here you will be able to see the list of red team attacks with statuses Pending, Running, and Completed#

4. Let’s create a new Attack by clicking on Create Attack.#

5. Type a Name for the attack that reflects the scenario or risk area you want to test.#

6. In the red teaming tab, choose the Attack Type you want to run#

7. From the list of available Probes, select the probe set that matches your chosen attack type or scenario.#

8. Select the Converter that defines how prompts and responses will be transformed or formatted before being sent to the model.#

9. Choose the Scorer that will evaluate the assistant’s responses.#

10. Review your configuration: Attack Type, Probe, Converter, and Scorer to ensure they match your testing goals.#

11. Click on the start attack button to start the red ream attack.#

12. A new job will appear in the jobs table with status Pending. After a short time, the status will update to Running as the attack is executed.#

13. Click on the Refresh Table button to update the job list and check for progress.#

14. Once the attack completes, the job status will change to Completed.#

15. Click on the completed job to open the detailed results view.#

16. Here you can inspect the prompts, model responses, and Scorer outputs for each red team example.#

17. You can also download the complete html report to view the vulnerable prompts (if any).#

1. For the Assistant you want to red team, click on the three dots next to it.

2. Click on Red Team to open the red teaming tab.

3. Here you will be able to see the list of red team attacks with statuses Pending, Running, and Completed

4. Let’s create a new Attack by clicking on Create Attack.

5. Type a Name for the attack that reflects the scenario or risk area you want to test.

6. In the red teaming tab, choose the Attack Type you want to run

7. From the list of available Probes, select the probe set that matches your chosen attack type or scenario.

8. Select the Converter that defines how prompts and responses will be transformed or formatted before being sent to the model.

9. Choose the Scorer that will evaluate the assistant’s responses.

10. Review your configuration: Attack Type, Probe, Converter, and Scorer to ensure they match your testing goals.

11. Click on the start attack button to start the red ream attack.

12. A new job will appear in the jobs table with status Pending. After a short time, the status will update to Running as the attack is executed.

13. Click on the Refresh Table button to update the job list and check for progress.

14. Once the attack completes, the job status will change to Completed.

15. Click on the completed job to open the detailed results view.

16. Here you can inspect the prompts, model responses, and Scorer outputs for each red team example.

17. You can also download the complete html report to view the vulnerable prompts (if any).