How It Works
From payload generation to security evaluation — here's how InvisiblePrompt helps you test AI systems against prompt injection attacks.
Step 1~1 min
Configure Payload
Select an encoding method for your invisible prompt injection payload.
- →Choose from 4 encoding methods: zero-width characters, homoglyph substitution, whitespace encoding, or HTML hiding
- →Customize the payload text and target injection context
- →Preview the encoded output before deployment
Step 2~2 min
Set Target
Connect InvisiblePrompt to the AI system you want to test.
- →Configure local providers (Ollama, OpenAI-compatible endpoints) or cloud providers (OpenAI, Anthropic)
- →Set API endpoints and authentication credentials
- →Define test parameters: model, temperature, max tokens
Step 3~30 sec
Inject Payload
Deploy the invisible payload into the target system's input field.
- →Payload is injected directly into DOM input fields via the Chrome extension
- →Invisible characters are undetectable to the AI system's preprocessing
- →Supports batch injection across multiple input fields simultaneously
Step 4~1 min
Analyze Response
Review how the AI system processed the injected payload.
- →InvisiblePrompt captures the AI system's raw response for analysis
- →Detect deviations from expected behavior, instruction overrides, or data leaks
- →Side-by-side comparison with a control (non-injected) response
Step 5~30 sec
Evaluate Results
Get a comprehensive security assessment with actionable recommendations.
- →Three evaluation categories: Bypass Detected, Partial Compromise, No Effect
- →Severity scoring with clear pass/fail thresholds
- →Export results as JSON or CSV for integration with your security reporting pipeline