Chatbot Evaluation

Red-team live chatbot websites using automated browser control.

Chatbot Evaluation (Website Mode) allows you to evaluate live chatbot websites through end-to-end red-team testing. Instead of calling an API directly, Know Your AI uses a browser control agent to open your website in the cloud, interact with your chatbot widget just like a real user would, and capture both the responses and screenshots throughout the process.

How it works

1. Provide your website URL

   In your product settings, configure the website URL where your chatbot is deployed, along with the CSS selectors for the chat input field and response area.

2. Select datasets

   Choose attack prompts from the Dataset Marketplace or use your own custom datasets.

3. Launch the evaluation

   Know Your AI spins up a cloud-based browser environment and deploys a browser control agent to automate the interaction with your chatbot.

4. Automated interaction

   The browser control agent navigates to your website, locates the chatbot widget, types each attack prompt, waits for the response, and captures screenshots at every step.

5. Judge responses

   Each captured response is passed to the judgment model, which scores it for vulnerabilities, producing isVulnerable, confidenceScore, and judgeAnalysis.

6. Review results

   View per-prompt results, screenshots, compliance reports, and an overall security score.

When to use Chatbot Evaluation

Chatbot Evaluation is ideal when:

• Your AI is deployed as a chatbot widget embedded on a website
• You want to test the full end-to-end user experience, including UI behavior
• You need to evaluate how your chatbot handles attacks in a real browser environment
• You want to capture visual evidence (screenshots) of chatbot responses
• Your chatbot requires authentication or multi-step interactions before testing

Browser control agent

Know Your AI uses an intelligent browser control agent to automate chatbot interactions in the cloud:

• Cloud-based browser — a full browser instance runs in a secure cloud environment
• Automated navigation — the agent automatically navigates to your website and locates the chatbot widget
• Smart interaction — the agent types prompts, clicks buttons, and waits for responses just like a real user
• Login detection — the system automatically detects if your website requires authentication and can pause for manual login via a live viewer
• Live viewer — watch the browser session in real time through an embedded live preview
• Screenshot capture — the agent captures screenshots at every step of the interaction for visual evidence and audit trails

Website connection configuration

To run a Chatbot Evaluation, your product must be configured as a Website product type with:

• Website URL — the URL where your chatbot is deployed
• Input selector — CSS selector for the chat input field
• Response selector — CSS selector for the chatbot’s response area
• Submit selector — CSS selector for the send/submit button (optional)

Evaluation pipeline

Navigate to Website → Locate Chatbot → Type Prompt → Capture Response & Screenshot → Judge → Store Results

For each prompt in the selected datasets:

1. The browser control agent navigates to your website (or reuses the active session)
2. The agent locates the chat input using your configured selector
3. The attack prompt is typed into the input field and submitted
4. The agent waits for the chatbot to respond and captures the response text
5. A screenshot is taken of the current browser state
6. The judgment model evaluates the prompt-response pair
7. Results, screenshots, and compliance analysis are stored

Screenshot library

Every Chatbot Evaluation automatically captures screenshots throughout the testing process. These screenshots are stored in the Screenshot Library, which provides:

• Visual audit trail — browse all captured screenshots organized by evaluation run and prompt
• Per-prompt screenshots — view the exact browser state when each attack prompt was tested
• Full-resolution images — screenshots are stored in S3 and accessible via presigned URLs
• Pagination — efficiently browse through large collections of screenshots
• Evidence for compliance — screenshots serve as visual evidence for compliance reporting and audit purposes

Access the Screenshot Library from the Library tab on any Chatbot Evaluation run page.

Results & insights

After a Chatbot Evaluation run completes, you get:

• Security score — overall vulnerability percentage across all tested prompts
• Per-prompt results — individual pass/fail verdicts with detailed judge analysis
• Screenshots — visual evidence of every chatbot interaction during the evaluation
• Compliance report — automated CCPA/CPRA violation analysis with evidence
• Live evaluation replay — review the browser session through the live viewer
• Real-time console — streaming execution logs showing each prompt, response, and judgment
• Run history — all past runs are stored for comparison and trend analysis

Scheduling

Like Model Evaluations, Chatbot Evaluations can be scheduled:

• Hourly, daily, weekly, or monthly intervals
• Custom cron expressions for fine-grained control
• Enable or disable schedules at any time

Scheduled chatbot evaluations are especially valuable for monitoring deployed chatbots and detecting regressions caused by model updates or configuration changes.

Related docs

Model Evaluation

Red-team your AI model via direct API testing.

Datasets

Browse attack datasets and upload your own.