NowAssist Guardian

[abdulshaafi]

I have enabled the Offensiveness and Prompt Injection guardian in Now Assist Guardian for ITSM. The guardian are working correctly during incident summary generation when the incident is created by me. If offensive content is added to the incident description, Now Assist blocks the summary as expected.

However, when another user creates an incident with the same short description and offensive content in the description, and I open that incident and generate the summary, Now Assist generates the summary instead of blocking it.

How to configure it correctly, any help would be appreciated

[KPNow]

Hi @abdulshaafi :

This behavior probably occurs because Now Assist Guardian might be differentiating between active user input and contextual record data. When you directly create the incident or enter text, Guardian treats that text as your active prompt input and immediately evaluates it against your Offensiveness and Prompt Injection guardrails. However, when you trigger the summary on an existing incident created by someone else, that offensive description is fetched as background context data (e.g., from sys_id / description fields) rather than treated as live prompt input. If Guardian is only set to evaluate user input or if the output generated by the LLM sanitizes the tone enough to fall below the output threshold, the summary bypasses the input filter and completes successfully.

To resolve this and ensure consistent blocking across all records, navigate to Now Assist Admin > Settings > Now Assist Guardian (or open the Incident Summarization skill configuration in Now Assist Skill Kit). Verify that Output Screening is active alongside input screening and set to Block and Log for the Incident Summarization skill. Additionally, recommend reviewing the prompt template in Skill Kit to ensure record context variables (such as description and work_notes) are flagged for Guardian evaluation, and try lowering the Guardian severity threshold (e.g., setting it to Low severity) so that any summary generated from toxic source context is intercepted before reaching the UI.