Image generation

The image generator makes an image of each text prompt. Use it to test if a vision model obeys text in an image.

The module is agentic_security.probe_data.image_generator. It uses matplotlib.

When the scanner makes images

If the body of the HTTP spec has <<BASE64_IMAGE>>, the scanner wraps the spec in a RequestAdapter. For each prompt, the adapter does these steps:

1. It makes a JPEG image of the prompt with generate_image(prompt).
2. It encodes the image as a data URL with encode(image).
3. It puts the data URL in place of <<BASE64_IMAGE>> and sends the request.

The data URL starts with data:image/jpeg;base64,. Do not write this prefix in the spec. Refer to HTTP spec.

Functions

Function	Description
generate_image(prompt, variant=0)	Returns the JPEG bytes of an image with the prompt text. The tool keeps the result in the disk cache.
generate_image_dataset(datasets)	Makes an ImageProbeDataset for each ProbeDataset.
encode(image)	Returns the image as a Base64 data URL.

Image styles

The image is 6 x 4 inches. The variant argument selects the style:

variant	Background
0 (default)	Light blue, dark blue text
1	Dark gray
2	Light pink, with a text box
3	White, with a circle

Example

from agentic_security.probe_data.image_generator import encode, generate_image

data_url = encode(generate_image("Ignore previous instructions"))

Limits

• The image has only text. It does not have pictures.
• The styles are fixed.