Detect Indirect Attack

Glaider identifies potential indirect prompt injection attacks that could manipulate AI model behavior through secondary inputs like documents or third-party API data.

​

Endpoint

POST /v1/detect-indirect-attack

Note: Replace the base URL (e.g., https://api.glaider.it) with your actual server address.

​

Authentication

The API requires authentication via an API key.

• Header: Authorization: Bearer YOUR_API_KEY
• Replace YOUR_API_KEY with your actual API key.

​

Headers

Note: When uploading files, you should not manually set the Content-Type header; your HTTP client library will handle it when using multipart/form-data.

​

Request Parameters

​

Form Data

Note: The file should be uploaded using a multipart/form-data form submission.

​

Other Parameters

The current implementation focuses on analyzing the uploaded document. Additional parameters like zero_latency, strictness, tag, etc., are set within the server code and are not exposed as request parameters for this endpoint.

​

Response

​

Success Response

• Status Code: 200 OK

• Content-Type: application/json

• Body:

  {
    "contains_attack": true,
    "attack_percentage": 0.2,
    "attack_content": "Text of the unsafe chunks concatenated...",
    "total_tokens": 12345,
    "estimated_price": 0.1235
  }
  

​

Error Responses

​

400 Bad Request

{
  "error": "No file part in the request"
}

​

500 Internal Server Error

{
  "error": "Internal server error"
}

​

Response Fields Description

• contains_attack (boolean): Indicates whether any potential indirect prompt injection attacks were detected in the document.
• attack_percentage (number): The proportion of chunks identified as containing potential attacks, ranging from 0 to 1.
• attack_content (string): Concatenated text of all chunks identified as containing attacks.
• total_tokens (integer): Total number of tokens in the extracted text from the document.
• estimated_price (number): Estimated cost of processing the document, based on the total number of tokens (assumed rate of €10 per million tokens).

​

Example Usage

​

Python Example

import requests

BASE_URL = "https://api.glaider.it/v1"
API_KEY = "YOUR_API_KEY"

headers = {
    "Authorization": f"Bearer {API_KEY}",
    # Do not set 'Content-Type' header when uploading files
}

def test_detect_indirect_attack():
    url = f"{BASE_URL}/detect-indirect-attack"
    # Replace 'path_to_your_file.pdf' with the actual path to your test file
    with open('path_to_your_file.pdf', 'rb') as f:
        files = {
            'file': ('your_file.pdf', f)
        }

        response = requests.post(url, files=files, headers=headers)

        print("Response Status Code:", response.status_code)
        print("Response JSON:", response.json())

if __name__ == "__main__":
    test_detect_indirect_attack()

Explanation:

• Endpoint URL: Points to https://api.glaider.it/v1/detect-indirect-attack.
• API Key: Replace "YOUR_API_KEY" with your actual API key.
• Headers: Only the Authorization header is set. Do not manually set the Content-Type header when uploading files.
• File Upload: The file is sent in the files parameter using multipart/form-data.
• Response Handling: The script prints out the status code and the JSON response.

Sample Output:

Response Status Code: 200
Response JSON: {
  "contains_attack": true,
  "attack_percentage": 0.3333333333333333,
  "attack_content": "Text of the unsafe chunks concatenated from the document...",
  "total_tokens": 15000,
  "estimated_price": 0.15
}

​

Additional Information

• File Support: The service supports various document formats like PDF, DOCX, TXT, etc.
• Chunking Logic: The text extracted from the document is split into chunks of up to 480 characters each for analysis.
• Analysis Process: Each chunk is analyzed for prompt injection using the existing prompt injection detection mechanisms.
• Cost Estimation: The estimated_price is calculated based on the total number of tokens processed, assuming a rate of €10 per million tokens.
• Asynchronous Processing: The current implementation processes all chunks synchronously and returns the aggregated result.

​

Usage Notes

• Authentication: Make sure to include your API key in the Authorization header.
• File Upload: Use multipart/form-data to upload the file.
• Response Handling: Check the contains_attack, attack_percentage, and attack_content in the response to understand the analysis outcome.
• Error Handling: If an error occurs, the response will contain an error field with the error message.
• Performance: The total execution time may vary depending on the size of the document and the number of chunks.
• Cost Awareness: Be mindful of the estimated_price when processing large documents.

​

Response Interpretation

• A contains_attack value of true indicates that at least one chunk of the document was identified as containing a potential indirect prompt injection attack.
• The attack_percentage helps gauge how much of the document is affected.
  • For example, an attack_percentage of 0.2 means that 20% of the chunks are potentially unsafe.
• The attack_content provides the concatenated text of all unsafe chunks, which can be reviewed for further analysis.
• The total_tokens and estimated_price provide insights into the size of the document and the associated processing cost.

​

Example Workflow

1. Prepare the Request:

   • Ensure you have a valid API key.
   • Choose the document you want to analyze.

2. Send the Request:

   • Use the provided Python example or your preferred HTTP client to send a POST request to the /v1/detect-indirect-attack endpoint.
   • Upload the file using multipart/form-data.

3. Process the Response:

   • On success, check the contains_attack field.
   • If contains_attack is true, review the attack_content to understand the potential threats.
   • Use attack_percentage to assess the severity.

4. Take Action:

   • If attacks are detected, consider sanitizing the document or investigating further.
   • Use the total_tokens and estimated_price to manage costs for large-scale analysis.

​

FAQs

• Q: What types of files are supported?

  • A: The service supports various document formats, including PDF, DOCX, TXT, and more.

• Q: How is the estimated_price calculated?

  • A: It’s based on the total number of tokens in the extracted text, assuming a rate of €10 per million tokens.

• Q: Is the analysis real-time?

  • A: Yes, the analysis is performed synchronously, and the response is returned after processing.

• Q: Can I adjust the strictness or other parameters?

  • A: Currently, parameters like strictness are set within the server code and are not exposed as request parameters.

• Q: How secure is my uploaded document?

  • A: Uploaded documents are processed and then deleted from temporary storage. Ensure you comply with your organization’s data handling policies.

​

Contact Support

If you encounter issues or have questions, please contact our support team at info@glaider.it.