Overview
The AI Guardrails Dashboard is a purpose-built monitoring surface within Netskope Advanced Analytics that gives security teams visibility into prompts and responses within GenAI apps across the enterprise and how well AI-specific security policies are performing.
Netskope One AI Guardrails is a runtime defense layer that analyzes every prompt submitted to and every response received from AI applications in real time. It detects and blocks adversarial techniques such as prompt injection and jailbreaking, moderates harmful or discriminatory content (hate speech, crimes, weapons), prevents IP and copyright exposure, and maps each detection to established industry frameworks — MITRE ATLAS and OWASP Top 10 for LLMs — so that security operations teams have standardized, actionable context.
The dashboard aggregates these detections and enforcement actions into a structured, filterable view. It is the primary operational interface for practitioners who need to:
- Monitor the volume and nature of AI interactions subject to guardrail policies
- Understand the balance of blocked versus monitored (allowed) traffic
- Identify high-risk content categories, matched keywords, and vulnerable applications
- Align findings to MITRE ATLAS techniques and OWASP LLM risks for SecOps workflows
- Investigate specific users or applications generating the most policy events
All dashboard widgets are based on the ‘Alerts’ data collection and share a common set of global filters that allow analysts to scope findings by date range, activity type, content category, application, user group, and individual user.
Data Attributes
To support custom reporting and tailored analysis, the following fields are available for the AI Guardrails alert schema in Advanced Analytics.
| ATTRIBUTE | DESCRIPTION |
|---|---|
| Profile Name | The guardrail profile applied at time of detection |
| Matched Category Name | The risk category triggered by the interaction |
| Matched Category Level | Severity level of the matched category |
| Matched Prompt/Response | The prompt or response that triggered detection |
| Matched Prompt/Response Level | Severity level of the matched prompt or response |
| Matched Keyword | Specific keyword(s) that triggered detection |
| OWASP Top 10 LLM | Mapped OWASP LLM risk classification |
| MITRE ATLAS | Mapped MITRE ATLAS technique |
| Matched Content | Full content of the matched interaction |
These attributes can be used to build custom widgets and dashboards, apply targeted filters, or enrich existing views with AI-specific threat context.
Global Dashboard Filters
All widgets on the dashboard contain the following filters.
| DISPLAY NAME | DATA TYPE | DESCRIPTION |
|---|---|---|
| Event Date | Date Filter | Scopes all tiles to a rolling or fixed time window. Default value is the last 30 days. Supports relative timeframes (e.g., last 7 days, last 90 days) via the relative timeframes UI control. |
| Activity | String Filter | Filters by the direction of the AI interaction — specifically whether the event was generated by a user Prompt (Post) or an AI model Response. Enables analysts to isolate prompt-side versus response-side risk. |
| Category | String Filter | Filters by the application category (e.g., Generative AI, Productivity). Useful for narrowing analysis to a specific segment of the app estate. |
| Application | String Filter | Filters by specific application name (e.g., ChatGPT, Google Gemini, Microsoft Copilot). Enables per-app policy review and compliance reporting. |
| User Group | String Filter | Filters by organizational user group or directory group. Allows targeted analysis for departments, business units, or risk cohorts. |
| User | String Filter | Filters by individual user identity (typically email address). Enables user-level investigation workflows for incident response. |
Dashboard Widget Reference
The dashboard starts with a high-level overview of the prompts and responses detected in your environment, which helps you understand how heavily your users are querying GenAI apps and how much traffic is blocked/allowed. This section also provides visibility into top prompt categories detected and the matched key words, which helps you uncover risky prompts and better fine tune your policies to control these prompts.
The following sections describe each widget in the order they appear on the dashboard, covering visualization type, business purpose, underlying logic, and the NAA – AI Guardrails Schema fields used.
Total Prompt/Response Detected
Description
This widget displays the total count of distinct transactions (prompts and responses combined) that were inspected and flagged by the AI Guardrails engine during the selected time period. It is the primary volume indicator on the dashboard and establishes the baseline scale of AI guardrail activity across the organization.
Business Use Case
Security leaders and operations teams use this KPI to understand overall AI usage exposure at a glance. A rising count may indicate expanding AI adoption, new policy triggers, or emerging attack patterns. It serves as the starting point for any guardrails-related investigation and is frequently used in executive reporting to communicate the scope of AI security monitoring.
Total Blocked Widget
Description
This widget shows the count of distinct transactions where the guardrails policy enforcement resulted in an active block action — meaning the prompt or response was prevented from reaching its destination. It is a direct indicator of how many interactions were stopped by policy rather than simply logged.
Business Use Case
The Total Blocked KPI is a core policy efficacy metric. A high blocked count relative to total detections signals that policies are set to an active enforcement posture rather than monitoring-only mode. Security teams use this number to validate that real-time protection is functioning, to report on enforcement effectiveness to compliance stakeholders, and to identify periods of elevated adversarial activity. A sudden spike in blocks may indicate a coordinated jailbreak or prompt injection campaign targeting the organization’s AI tools.
Allowed (Monitored)
Description
This widget counts all detected transactions where the guardrails engine flagged activity but the policy action was not a block — meaning the interaction was allowed to proceed while still being logged and monitored. This reflects an ‘alert’ or ‘allow’ posture rather than active enforcement.
Business Use Case
The Allowed (Monitored) KPI complements the Total Blocked widget and together they reflect the organization’s policy posture split. A high allowed count compared to blocked suggests the organization is primarily in a detection and monitoring mode, which is typical in early deployment phases or for lower-risk content categories. Security teams use this metric to evaluate whether current policies should be escalated from alert to block based on observed risk patterns. It also supports capacity planning by showing how much AI traffic is flowing through the inspection layer without being stopped.
Prompt vs Response
Description
This pie chart breaks down the total detected transactions into two segments: those triggered by user Prompts (Post activity) and those triggered by AI model Responses. Each slice displays the percentage of total events it represents. Based on sample dashboard data, prompts account for approximately 75% of detections and responses approximately 25%.
Business Use Case
Understanding whether risk originates predominantly from user behavior (prompts) or from AI model outputs (responses) fundamentally shapes the remediation strategy. A prompt-heavy distribution indicates that users are the primary source of policy violations — whether through accidental sensitive data sharing or deliberate jailbreak attempts. A response-heavy distribution points to problems with the AI model itself generating harmful, biased, or inappropriate content. This chart helps security architects decide whether to tighten input controls, output filtering, or both.
Category Description Widget
Description
The Category Distribution treemap visualizes the relative frequency of content categories that triggered guardrail detections. Each rectangle’s size is proportional to the number of distinct transactions associated with that matched category. Categories visible in sample data include Weapons, Prompt Injection and Jailbreaking, Crimes, and Hate Speech, among others.
Business Use Case
This widget is the primary tool for understanding the risk landscape composition of AI traffic within the organization. It immediately reveals which types of harmful or policy-violating content are most prevalent, allowing security teams to prioritize policy tuning, user awareness training, and escalation thresholds. For example, a large Prompt Injection and Jailbreaking block indicates active adversarial attempts on the organization’s AI tools, warranting immediate investigation. A large Hate Speech block may indicate cultural issues requiring HR engagement alongside technical controls.
Top Keywords Matched Widget
Description
This horizontal bar chart ranks the most frequently matched keywords or terms that triggered guardrail policy detections, sorted in descending order by transaction count. Each bar represents a specific keyword or phrase and its length indicates how many distinct transactions contained that term. Sample data shows keywords such as color terms (Green, green, Blue, blue) alongside null values, suggesting pattern-matching rules that extend beyond literal dangerous terms.
Business Use Case
Keyword analysis is critical for validating and tuning content inspection policies. If benign words such as color names appear prominently, it indicates that policy keyword lists may need refinement to reduce false positives. Conversely, if highly specific harmful terms appear frequently, it may indicate coordinated misuse. Security administrators use this widget to review and optimize the keyword match rules within AI Guardrails profiles, ensuring that policies are both effective and precise. It also helps identify emerging language patterns used in jailbreak or social engineering attempts.
Security Frameworks – MITRE ATLAS Widget
Description
This column chart maps guardrail detections to specific MITRE ATLAS (Adversarial Threat Landscape for Artificial-Intelligence Systems) techniques. Each column represents a MITRE ATLAS technique ID and its height shows the number of distinct transactions linked to that technique. Sample data shows techniques such as AML.T0048, AML.T0051.000, AML.T0054, and AML.T0051.001, representing techniques related to prompt injection, model inversion, and adversarial inputs.
Business Use Case
MITRE ATLAS alignment is a cornerstone of enterprise AI security operations. This widget allows security operations center (SOC) analysts and threat intelligence teams to speak a common language with the broader security community when describing AI-specific attacks. By mapping detections to ATLAS techniques, teams can correlate AI guardrail findings with threat intelligence feeds, prioritize investigation based on technique severity and novelty, and produce compliance-ready reporting. It also enables benchmarking against industry peers and integrating AI security findings into existing SIEM and SOAR workflows that already leverage MITRE frameworks.
Security Frameworks – OWASP Top 10 LLM Widget
Description
This column chart maps guardrail detections to the OWASP Top 10 for Large Language Model Applications risk categories. Each column represents an OWASP LLM risk identifier and its height reflects the transaction count. Sample data shows LLM09:2025 Misinformation and LLM01:2025 Prompt Injection as the top risks, with a null category also present. The widget includes a developer note indicating that the OWASP risk values arrive as comma-separated strings from the UI and require format validation.
Business Use Case
The OWASP Top 10 for LLMs is one of the most widely referenced AI security frameworks for application development and enterprise AI risk management. This widget directly connects operational detection data to a globally recognized risk taxonomy, enabling security and compliance teams to assess which OWASP risks are most prevalent in the organization’s AI usage. It supports risk register maintenance, board-level reporting, and regulatory compliance efforts. SOC teams also use this widget alongside the MITRE ATLAS chart to get a complete picture of both attacker techniques (ATLAS) and the resulting application-layer risks (OWASP).
Allowed vs Blocked Widget
Description
This full-width visualization uses the ‘donut multiples’ chart type to display the distribution of enforcement actions (allowed versus blocked) broken out by activity type (Prompt vs Response). Each donut represents one action category and is subdivided by activity, giving a simultaneous view of policy posture across both prompt and response directions. The accompanying note reads: ‘Action distribution for detected prompts and responses.’
Business Use Case
The Allowed vs Blocked widget provides a nuanced policy posture view that goes beyond the top-row KPI cards. By breaking out the block/allow ratio per activity type, it reveals whether blocking is applied consistently to both prompts and responses or concentrated in one direction. For example, an organization might block all jailbreak attempts in prompts but only monitor (allow) harmful content in responses during a policy tuning phase. This granularity helps policy administrators identify asymmetric enforcement gaps and prioritize where to tighten controls. It is also a key artifact for compliance reviews that require documented evidence of enforced versus monitored policy actions.
Top Applications by Policy Action Taken Widget
Description
This stacked column chart shows the top 10 applications generating AI Security guardrail events, with each column broken into stacked segments representing the different enforcement actions taken (block, alert, allow). This allows direct comparison of both total event volume and action distribution across applications. Sample data shows ChatGPT with 13 detections and Google Gemini with 3.
Business Use Case
Application-level visibility is essential for risk-based AI governance. This widget answers the question: which AI applications are generating the most policy events, and are they being blocked or merely monitored? High-volume applications with a large ‘allowed’ segment may need stricter policies, while applications with predominantly blocked traffic may be candidates for user access reviews or complete policy blocks. Security teams also use this widget to compare the risk profile of sanctioned AI tools (e.g., Microsoft Copilot) against unsanctioned applications (shadow AI), and to build the business case for application-specific guardrail policies.
Top Users by Policy Action Taken
Description
This horizontal bar chart lists the top 10 users generating AI Security guardrail events, with each bar segmented by enforcement action taken. A row total is included in the underlying data (though hidden in the chart display). Users are ranked by total event count in descending order. Sample data shows ayang@netskope.com as the top user with 16 total detections, all classified as Detection action type.
Business Use Case
User-level attribution is critical for both incident response and insider risk programs. This widget surfaces the individuals who interact most frequently with AI systems in ways that trigger security policies. High-volume users warrant investigation to determine whether activity is benign (e.g., a power user working with AI tools legitimately) or malicious (e.g., exfiltration attempts or persistent jailbreak campaigns). The action breakdown per user also reveals whether specific individuals are subject to active blocks versus being monitored — a key distinction for escalation decisions. This widget is also used in HR-engaged risk reviews, security awareness targeting, and access policy enforcement.
NAA – AI Guardrails Schema Reference
The following table lists the key fields for the AI Guardrails schema.
| PROFILE NAME (NAME OF PROFILE THAT MATCHED THE POLICY) | DATA TYPE | DESCRIPTION |
|---|---|---|
| Matched Category | Dimension | Matched predefined categories that triggered an alert. |
| Matched Prompt/Response | Dimension | Matched prompt or response that triggered an alert. |
| Matched Keyword | Dimension | Matched keyword that triggered an alert. |
| OWASP top10 LLM | Dimension | The standard framework for identifying the most critical security vulnerabilities in Generative AI systems. |
| MITRE ATLAS | Dimension | Adversarial Threat Landscape for Artificial Intelligence Systems is a knowledge base of adversary tactics and techniques based on real-world attacks and theoretical vulnerabilities in Artificial Intelligence (AI) and Machine Learning (ML) systems. |
Appendix — Framework Reference
MITRE ATLAS Techniques Referenced
The following MITRE ATLAS technique IDs appear in dashboard data and are documented here for analyst reference:
| NAME | DATA TYPE | DESCRIPTION |
|---|---|---|
| Backdoor ML Model | Technique | Adversarial manipulation of ML model weights or training data to create covert backdoors that can be triggered by specific inputs. |
| LLM Prompt Injection | Sub-technique | Malicious content embedded in user prompts that overrides the LLM's system instructions, causing it to behave outside its intended boundaries. |
| Indirect Prompt Injection | Sub-technique | Adversarial instructions injected via external content (web pages, documents, emails) that the LLM processes, rather than directly in the user prompt. |
| LLM Jailbreak | Technique | Techniques designed to bypass an LLM's built-in safety guardrails, forcing it to produce outputs it is programmed to refuse. |
OWASP Top 10 LLM Risks Referenced
The following OWASP Top 10 for LLM Applications (2025) risk categories appear in dashboard data:
| NAME | DATA TYPE | DESCRIPTION |
|---|---|---|
| Prompt Injection | Critical Risk | Adversarial inputs that manipulate LLM behavior by overriding system prompts or injecting unauthorized instructions into the model's context window. |
| Misinformation | High Risk | LLMs generating false, misleading, or hallucinated content that is presented as factual, leading to downstream harm, compliance issues, or reputational damage. |