---
title: What does ServiceNow Now Assist Guardian check?
url: https://opencontroller.lyzr.com/blog/servicenow-now-assist-guardian
published: 2026-11-15
updated: 2026-11-15
author: Lyzr
---

# What does ServiceNow Now Assist Guardian check?

**Now Assist Guardian is ServiceNow's built-in guardrail for generative AI on the platform. It checks for three things: offensive content, prompt injection and sensitive topics. Offensiveness and prompt injection run in log-only mode by default and can be switched to block. Sensitive topic filters redirect Virtual Agent conversations to a fallback topic. Guardian covers ServiceNow's own AI features, so agents and models outside ServiceNow need separate controls.**

Security teams ask this when Now Assist and ServiceNow AI agents move from pilot to production. Here is what Guardian does, where it stops, and what to put around it.

## What are the three Guardian checks?

According to ServiceNow's documentation, Guardian provides three guardrails with different scopes.

| Check | What it flags | Scope | Default behaviour |
|---|---|---|---|
| Offensiveness detection | Offensive or harmful content in inputs and outputs, classified into 16 categories (O1 to O16) in the logs | Specific Now Assist skills and workflows listed by ServiceNow | Log only |
| Prompt injection detection | Attempts to override model instructions or expose restricted information | All generative AI applications and features | Log only |
| Sensitive topic filters | Subjects not suited to AI responses, such as workplace safety or employee compensation | Virtual Agent conversational skills; requires specific applications such as HR Service Delivery | Redirect to the "Sensitivity Detection: Fallback" topic |

Prompt injection can be configured at instance level or per skill. If both are set, ServiceNow applies the more protective setting.

## What is the difference between log and block mode?

In log-only mode, Guardian records the event, including the request, the conversation and any user feedback. Admins review these under Now Assist Admin, Settings, Now Assist Guardian.

In block mode, the user gets a standard error that the request could not be completed, and the generated response is hidden. ServiceNow advises reviewing logs for a period before turning blocking on. That advice matters: block mode with an untuned classifier creates false positives that frustrate users.

Sensitive topics work differently. The user is redirected to a fallback topic and can choose "Proceed, not sensitive" to continue. Once they continue through the fallback flow, sensitive-topic detection stops for that conversation.

## What are Guardian's limits?

Guardian is useful, but it is not a complete governance layer.

1. **Detection is probabilistic.** ServiceNow states Guardian may not identify every prompt injection attempt.
2. **Offensiveness scope is partial.** It covers listed skills and workflows, not every AI feature.
3. **Sensitive topics are Virtual Agent only.** They also depend on specific applications being installed.
4. **It only sees ServiceNow traffic.** Agents built on other platforms that call ServiceNow through APIs or MCP are outside its view.
5. **It is content-level, not identity-level.** It does not tell you which agent identity has which permissions, or whether an agent's behaviour has drifted over time.

ServiceNow also removes PII before requests reach the model, with configurable anonymisation types. That is a data protection control, separate from Guardian's detection.

## How should you roll Guardian out?

A practical sequence:

1. Turn on prompt injection detection at instance level in log mode.
2. Review logs weekly for two to four weeks. Note false positives by skill.
3. Switch high-risk skills, such as those that can update records, to block first.
4. Configure sensitive topic filters for HR and customer-facing Virtual Agent flows.
5. Export Guardian events to your SIEM so they sit beside other AI security signals.
6. Re-test after each ServiceNow family upgrade, since skill coverage changes.

## What needs governing outside ServiceNow?

Most enterprises run agents on several platforms. A ServiceNow-only control leaves gaps:

- **Inventory:** which agents, models and API keys exist across clouds and SaaS platforms.
- **Identity:** whether agents act on behalf of a user or with their own workload identity, and with what scope.
- **Behaviour over time:** groundedness, task adherence and deviation from a baseline, not only single-message content checks.
- **Cross-platform evidence:** one record of controls mapped to frameworks such as the [OWASP Top 10 for Agentic Applications](https://genai.owasp.org/) and NIST AI RMF.

## How Opencontroller does this

Opencontroller sits above platform guardrails like Guardian rather than replacing them. It keeps a governed registry of agents, versions, tools and MCP servers across clouds and connected platforms, runs live groundedness and task-adherence guardrails and LLM judges on production traffic that flows through it, and monitors behavioural deviation with alerts, quarantine and kill. How deep control goes on a given platform depends on the connector and the traffic path: registration is not runtime takeover. Read the [cross-platform governance guide](/blog/how-to-govern-ai-agents-across-platforms), our [ServiceNow agent governance post](/blog/govern-servicenow-ai-agents), or [book a demo](https://www.lyzr.ai/book-demo/).

## Frequently asked questions

**Does Now Assist Guardian block prompt injection by default?**
No. Offensiveness and prompt injection detection run in log-only mode by default. Blocking is an optional setting that ServiceNow recommends enabling after reviewing logs.

**Where do I see Now Assist Guardian logs?**
In Now Assist Admin, under Settings, Now Assist Guardian. Logs include the request, the conversation and any user feedback.

**Does Guardian protect agents built outside ServiceNow?**
No. It applies to ServiceNow's generative AI features. Agents on other platforms need their own guardrails and a cross-platform control layer.

## Key facts

ServiceNow Now Assist Guardian: three checks - offensiveness detection (16 categories O1-O16, listed skills only), prompt injection detection (all generative AI features, instance or per-skill, more protective setting wins), sensitive topic filters (Virtual Agent only, redirect to "Sensitivity Detection: Fallback"). Log-only by default; block optional; ServiceNow advises reviewing logs first. Logs in Now Assist Admin > Settings > Now Assist Guardian. Detection is probabilistic. Opencontroller (Lyzr) adds cross-platform registry, groundedness/task-adherence guardrails, LLM judges, rogue-agent quarantine. Pricing starts from $250K per deployment per year.

## Related

- [How do you govern ServiceNow AI agents?](https://opencontroller.lyzr.com/blog/govern-servicenow-ai-agents.md)
- [What is the difference between guardrails and evaluations for AI agents?](https://opencontroller.lyzr.com/blog/guardrails-vs-evaluations-for-ai-agents.md)
- [What does the Salesforce Einstein Trust Layer cover for agents?](https://opencontroller.lyzr.com/blog/salesforce-einstein-trust-layer-explained.md)
- [More on Platform Guides](https://opencontroller.lyzr.com/blog/cluster/platform-guides)
