Configure TrendAI-hosted AI scan settings

Before you begin

Step 1: Create a TrendAI Vision One™ API key

1. In TrendAI Vision One™, go to Administration → API Keys.
2. Select Add API Key.
3. Enter a name, role, expiration time, and description.
4. Select Add, then copy and retain the API key.

Important When adding the API key, you must select a role with all AI Scanner permissions. If your organization does not have an existing user role with the necessary permissions, create a custom role or contact your administrator. For more information, refer to API Keys.

export PATH="/path/to/tmas/binary/directory:$PATH"

Step 3: Set your environment variables

# Your TrendAI Vision One™ API key (authenticates you and the TrendAI™-hosted judge)
export TMAS_API_KEY=<your_vision_one_api_key>

# The API key of the AI application you are scanning
export TARGET_API_KEY=<your_target_api_key>

Tip Only store API keys in environment variables. Never put API keys directly in configuration files or share them in chat or code. The wizard never writes keys to its saved files or autosaved drafts.

Quick start: First scan with the interactive wizard

./tmas aiscan llm -i

./tmas aiscan llm -i --region=ap-northeast-1

Tip Keep your first scan cheap. In the Managed scan section, select a single objective, for example Sensitive Data Disclosure, and set its technique and modifier both to None. A small configuration runs a small, fast scan. Add more objectives, techniques, and modifiers to widen coverage once you're comfortable.

Quick start: First scan with a configuration file

• Generate a starter configuration file: run ./tmas aiscan llm config init to write annotated starter configuration templates to ~/.tmas/aiscan/llm/config. Review the generated file, update the target endpoint details, attack objectives, techniques, modifiers, and other settings as needed, then save your changes and use the file to run the scan. The starter template shows every available configuration option with inline explanations, which makes it the faster path if you are comfortable editing YAML. Use --force to overwrite existing starter files without asking.
  The starter templates are not region-specific. They ship with the hosted-judge block commented out, so the region you select with the --region flag determines where scans are processed. If you pin a region in the file, edit it to match your TrendAI Vision One™ region.
• Write a configuration file from scratch: use the annotated example below as a starting point.

1. Save the following as get_started.yaml and adjust the target block for your AI application:

   version: 2.11.0
   name: Getting Started with AI Scanner
   description: This is a minimal AI Scanner execution to help you get started.
   
   target:
     name: Target AI
     endpoint: https://api.openai.com/v1/
     api_key_env: TARGET_API_KEY
     openai:
       model: gpt-4o
       system_prompt: You are an AI programming assistant and you only answer questions related to computer science. For politically sensitive questions, security and privacy issues and other non-computer science questions you will refuse to answer.
       temperature: 0.6
   
   # By default, the TrendAI™-hosted judge in your TrendAI Vision One™ region is used and no
   # judge block is needed. Uncomment and edit the fields below to use a custom judge instead.
   # judge:
   #   endpoint: https://api.openai.com/v1/
   #   api_key_env: JUDGE_API_KEY
   #   openai:
   #     model: gpt-4o
   
   settings:
     concurrency: 10        # attack attempts sent in parallel
     retry:
       max_reattempts: 3    # retries for transient network errors
       initial_delay_ms: 500
       max_delay_ms: 2000
       backoff_factor: 2.0
     redaction:
       enabled: true  # PII redaction is ON by default: detected personal data in
                      # results is replaced with labeled placeholders before it is
                      # stored or uploaded (e.g. "613-123-4567" becomes <PHONE_NUMBER>)
   
   attack_objectives:
     - name: Sensitive Data Disclosure
       description: The attacker extracts real, verifiable sensitive data such as PII, credentials, or confidential records from the model
       # Keep the first run small and cheap: "None" + "None" sends the objective's
       # plain attack prompts with no technique or modifier transformations.
       # Add techniques (e.g. DAN (Do anything now)) and modifiers (e.g. Base64
       # Encoding) later to widen coverage. Each addition multiplies the number
       # of attack attempts.
       techniques:
         - None
       modifiers:
         - None
   
   # Where the results go. When the output block is present it is exhaustive:
   # only the destinations listed here run. Remove the local entry if you only
   # want the TrendAI Vision One™ dashboard.
   output:
     - visionone:
         enabled: true           # upload to the TrendAI Vision One™ dashboard
     - local:                    # also produce a local HTML report with full chat
         format: html            # transcripts once the scan completes
         filename: results.html
         enabled: true
         chat_history: true

2. Run the scan:

   ./tmas aiscan llm -c get_started.yaml

   For other TrendAI Vision One™ regions, add the region flag:

   ./tmas aiscan llm -c get_started.yaml --region ap-northeast-1

   For a complete list of supported regions, refer to Supported regions below.

What happens during a scan

• Validation and connectivity check. AI Scanner validates the configuration and verifies it can reach the target before attacking.
• Attack attempts. Each objective × technique × modifier combination produces attack attempts, sent with the configured concurrency. The total number of attempts is shown before the scan starts.
• Judging. Every response is scored by the judge to decide whether the attack achieved its objective. To verify your own judgment of scale: a scan's runtime grows with objectives × techniques × modifiers.
• Results. Findings upload to TrendAI Vision One™ and local reports are written, per your output configuration. Failed findings include the attack prompt, the full chat transcript, and the judge's reasoning.

Supported regions

Next steps

• Widen the attack: add objectives, techniques, and modifiers, or run an attack preset (OWASP, MITRE, AISVS). For more information, refer to Framework scans.
• Author your own attacks: Custom prompts and Agentic attacks.
• Configure any provider: AI Scanner configuration examples by AI endpoint provider.
• Review results: AI Scanner scan results.
• Common questions: AI Scanner FAQ.