Reducing Guardian Model False PositivesIssue <!-- /*NS Branding Styles*/ --> .ns-kb-css-body-editor-container { p { font-size: 12pt; font-family: Lato; color: var(--now-color--text-primary, #000000); } span { font-size: 12pt; font-family: Lato; color: var(--now-color--text-primary, #000000); } h2 { font-size: 24pt; font-family: Lato; color: var(--now-color--text-primary, black); } h3 { font-size: 18pt; font-family: Lato; color: var(--now-color--text-primary, black); } h4 { font-size: 14pt; font-family: Lato; color: var(--now-color--text-primary, black); } a { font-size: 12pt; font-family: Lato; color: var(--now-color--link-primary, #00718F); } a:hover { font-size: 12pt; color: var(--now-color--link-primary, #024F69); } a:target { font-size: 12pt; color: var(--now-color--link-primary, #032D42); } a:visited { font-size: 12pt; color: var(--now-color--link-primary, #00718f); } ul { font-size: 12pt; font-family: Lato; } li { font-size: 12pt; font-family: Lato; } img { display: ; max-width: ; width: ; height: ; } } Content Moderation and Guardian Model may produce false positives, blocking legitimate requests in agentic workflowsSolution: Increase confidence thresholds from default 0.5 to higher values (up to 0.9)Impact: Reduces false positive rate (FPR) but may slightly decrease recallCaution: Do not exceed 0.8/0.9 threshold without testing, it might causes significant recall degradation. Facts<!-- /*NS Branding Styles*/ --> .ns-kb-css-body-editor-container { p { font-size: 12pt; font-family: Lato; color: var(--now-color--text-primary, #000000); } span { font-size: 12pt; font-family: Lato; color: var(--now-color--text-primary, #000000); } h2 { font-size: 24pt; font-family: Lato; color: var(--now-color--text-primary, black); } h3 { font-size: 18pt; font-family: Lato; color: var(--now-color--text-primary, black); } h4 { font-size: 14pt; font-family: Lato; color: var(--now-color--text-primary, black); } a { font-size: 12pt; font-family: Lato; color: var(--now-color--link-primary, #00718F); } a:hover { font-size: 12pt; color: var(--now-color--link-primary, #024F69); } a:target { font-size: 12pt; color: var(--now-color--link-primary, #032D42); } a:visited { font-size: 12pt; color: var(--now-color--link-primary, #00718f); } ul { font-size: 12pt; font-family: Lato; } li { font-size: 12pt; font-family: Lato; } img { display: ; max-width: ; width: ; height: ; } } Content moderation, the Guardian Model can still introduce false positives or false negatives. - Upcoming Improvements:We have two parallel workstreams to address this more comprehensively:- Prompt Optimization (Included in Generative AI Controller V13.0, with optimizations ongoing afterwards): We are refactoring the Guardian integration to reduce the amount of content sent to guardrail services, filtering out backend metadata (sys_ids, qualifiers, internal references) that cannot contain prompt injections or offensive content.This will significantly reduce false positives caused by complex agentic prompts- Model & Pipeline Enhancements (Ongoing): We continuously iterate on our Guardian model pipeline. The next versions will continue improving the handling of edge cases more gracefully and improve precision on enterprise agentic workflows Release<!-- /*NS Branding Styles*/ --> .ns-kb-css-body-editor-container { p { font-size: 12pt; font-family: Lato; color: var(--now-color--text-primary, #000000); } span { font-size: 12pt; font-family: Lato; color: var(--now-color--text-primary, #000000); } h2 { font-size: 24pt; font-family: Lato; color: var(--now-color--text-primary, black); } h3 { font-size: 18pt; font-family: Lato; color: var(--now-color--text-primary, black); } h4 { font-size: 14pt; font-family: Lato; color: var(--now-color--text-primary, black); } a { font-size: 12pt; font-family: Lato; color: var(--now-color--link-primary, #00718F); } a:hover { font-size: 12pt; color: var(--now-color--link-primary, #024F69); } a:target { font-size: 12pt; color: var(--now-color--link-primary, #032D42); } a:visited { font-size: 12pt; color: var(--now-color--link-primary, #00718f); } ul { font-size: 12pt; font-family: Lato; } li { font-size: 12pt; font-family: Lato; } img { display: ; max-width: ; width: ; height: ; } } Available with Zurich Patch 5 Cause<!-- /*NS Branding Styles*/ --> .ns-kb-css-body-editor-container { p { font-size: 12pt; font-family: Lato; color: var(--now-color--text-primary, #000000); } span { font-size: 12pt; font-family: Lato; color: var(--now-color--text-primary, #000000); } h2 { font-size: 24pt; font-family: Lato; color: var(--now-color--text-primary, black); } h3 { font-size: 18pt; font-family: Lato; color: var(--now-color--text-primary, black); } h4 { font-size: 14pt; font-family: Lato; color: var(--now-color--text-primary, black); } a { font-size: 12pt; font-family: Lato; color: var(--now-color--link-primary, #00718F); } a:hover { font-size: 12pt; color: var(--now-color--link-primary, #024F69); } a:target { font-size: 12pt; color: var(--now-color--link-primary, #032D42); } a:visited { font-size: 12pt; color: var(--now-color--link-primary, #00718f); } ul { font-size: 12pt; font-family: Lato; } li { font-size: 12pt; font-family: Lato; } img { display: ; max-width: ; width: ; height: ; } } The issue stems from how prompts are constructed in the agentic workflow and the limitations of the LLM models- User input can be minimal (e.g., "Request Service Change "), but the backend compiles this into an extremely long prompt containing hundreds of sys_ids, metadata attributes, qualifiers, and internal references- Guardian models can be sensitive to longer, complex contexts and may produce elevated confidence scores even when no actual privacy violation exists- The model can a score of 0.530—a borderline case just above the 0.5 threshold—triggering a false positive- Essentially: The model misinterpreted the presence of internal metadata (sys_ids, employee type filters, reference qualifiers) as potential PII exposure, when in reality this is standard ServiceNow catalog item data required for the workflow to function. Resolution<!-- /*NS Branding Styles*/ --> .ns-kb-css-body-editor-container { p { font-size: 12pt; font-family: Lato; color: var(--now-color--text-primary, #000000); } span { font-size: 12pt; font-family: Lato; color: var(--now-color--text-primary, #000000); } h2 { font-size: 24pt; font-family: Lato; color: var(--now-color--text-primary, black); } h3 { font-size: 18pt; font-family: Lato; color: var(--now-color--text-primary, black); } h4 { font-size: 14pt; font-family: Lato; color: var(--now-color--text-primary, black); } a { font-size: 12pt; font-family: Lato; color: var(--now-color--link-primary, #00718F); } a:hover { font-size: 12pt; color: var(--now-color--link-primary, #024F69); } a:target { font-size: 12pt; color: var(--now-color--link-primary, #032D42); } a:visited { font-size: 12pt; color: var(--now-color--link-primary, #00718f); } ul { font-size: 12pt; font-family: Lato; } li { font-size: 12pt; font-family: Lato; } img { display: ; max-width: ; width: ; height: ; } } 1) Go to flow designer 2) Choose actions and search for `execute moderation` and select it. 3) In Execute Moderations API action setting go to `REST step` and under `Request Content`, expand the `Request Body [Text]` Add the threshold like below return { "input": fd_data.action_inputs.input, "category_score_threshold": 0.9 }; And publish from the button on the top menu: NOTE: Now Assist Admin has a new feature starting with Australia Release, March 12,2026, Guarding model threshold can be handled through UI. Go to Now Assist Admin / Settings: