Views:

Evaluate the results of scans of your AI applications.

After a scan completes, its findings appear in AI Scanner in TrendAI Vision One™, where you can filter them, chart them over time, investigate individual attacks, and export reports.
To open the findings, go to AI Security → AI Application Security and open AI Scanner. Select Set up for instructions on configuring AI Scanner. For more information, refer to Configure TrendAI™-hosted AI scan settings or Configure self-hosted AI scan settings.

Results over time

For the selected target, the dashboard charts your scan results over time, so you can see whether a target is improving or regressing instead of comparing scans one by one:
  • Results over time: The distinct attacks that succeeded, broken down by severity. When multiple scans fall under a single data point, duplicate successful attacks are filtered down to distinct results.
  • Attack success rate over time: How often attacks succeed as your scans accumulate.
You can switch between per-day and per-scan views. Plot the aggregated per-day (or per-month) series, or show each individual scan to see how results vary from one run to the next.
Scan results are retained for 180 days, giving the over-time charts a wide view window. Every scan result carries the timestamp of when it was produced, shown in the Time column of the results table and in the finding details, so you can tell exactly when each result was recorded.

Filtering and exports

You can filter the results by Target, Selected scan, and Objective. Filters narrow every widget, the findings table, and exported reports at once. All objectives are selected by default, so nothing is hidden until you narrow the view.
Use the Export menu to download the results you're currently viewing as any of the following:
Format
What it contains
CSV
A spreadsheet of the findings table.
JSON
The results in structured form, for tooling and further analysis.
CycloneDX
The findings as a CycloneDX 1.6 report in Vulnerability Exploitability eXchange (VEX) form, for supply-chain tooling.
HTML
A self-contained, sortable report you can share or archive.
Executive report
A high-level summary of the overall security posture of the scanned target, suitable for leadership review.
All exports respect the filters currently applied. To bring results from an on-premises or air-gapped environment into TrendAI Vision One™, you can upload an exported HTML report back into the platform. Reports up to 100 MB are supported.

What the results show

The following elements are available after a scan completes:
Element
Description
Scan details
Basic information on your target AI model and the start time and duration of the scan.
Severity breakdown
An overview of the severity levels of successful attacks on your AI model, organized into Critical, High, Medium, and Low severity levels.
Top attack objectives
A chart displaying the most commonly used attack objectives targeting your AI model. An attack objective is the outcome that an attack on an AI model attempts to produce, such as the disclosure of sensitive data.
Top attack techniques
A chart displaying the most common techniques used to attack your AI model, such as DAN (Do Anything Now) prompts.
Framework compliance
A chart of the percentage of scans in which your model demonstrated compliance with global compliance standards, including the Open Worldwide Application Security Project (OWASP) Top 10 for Large Language Model Applications and MITRE ATLAS. For more information on scoring against the OWASP AI Security Verification Standard (AISVS), refer to Framework scans.
Attack outcome by objective
A chart of the outcomes of failed and successful attacks on your model, organized by attack objective.
Evaluation results
A table of the aggregated results of all scans of your AI model, including the scan settings, severity level, final attack outcome, and the time each result was produced.

Investigate an individual attack

Select View details in any row of the Evaluation results table to see the full evidence for that attack, including the following:
  • Original prompt: The initial prompt used to begin the chat.
  • Chat history: Select Chat history to view the full transcript of the chat between the user and the AI model (every turn that led to the outcome).
  • Attack objective and attack technique: The attack settings selected for the scan.
  • Framework compliance: The compliance status of the model against the frameworks the attack maps to.
  • Evaluation: The judge's reasoning for why the attack failed or succeeded.
Important
Important
Chat transcripts are included by default when the TrendAI™-hosted judge scores the scan. When you configure a custom judge, including the self-hosted service, transcripts are excluded by default to protect data privacy, and you can control inclusion per destination in the scan's output settings.

Local reports

In addition to TrendAI Vision One™, a scan can write reports to local files while it runs: JSON (the complete raw results), a JSON summary, Markdown, or a self-contained HTML report. Configure local reports in the interactive wizard's output settings or with the output block of the scan configuration file. For more information, refer to AI Scanner configuration examples by AI endpoint provider.