Probe data¶

The agentic_security.probe_data package loads the attack datasets. It can also change the prompts and make image or audio attacks.

Datasets¶

agentic_security.probe_data.data:

Function	Description
`prepare_prompts`	Loads and transforms the selected registry datasets for a scan.
`load_dataset_generic`	Loads a CSV URL or a Hugging Face dataset into a `ProbeDataset`.
`load_local_csv`	Loads all CSV files in `./datasets` into one `ProbeDataset`.
`load_local_csv_files`	Loads each CSV file in `./datasets` into a separate `ProbeDataset`.

ProbeDataset in agentic_security.probe_data.models keeps the prompts and the metadata of one dataset. ImageProbeDataset adds the images.

Transforms¶

agentic_security.probe_data.stenography_fn has encoding functions, for example rot13, base64_encode, and mirror_words. Refer to Stenography functions.

Images and audio¶

generate_image and generate_image_dataset in probe_data.image_generator. Refer to Image generation.
generate_audioform in probe_data.audio_generator. It makes a WAV file. On macOS it uses the say command. On other systems it uses gTTS.

from agentic_security.probe_data.audio_generator import generate_audioform

audio_bytes = generate_audioform("Hello, world!")

Prompt selection¶

probe_data.modules.rl_model has prompt selectors. Refer to RL prompt selection.

from agentic_security.probe_data.modules.rl_model import QLearningPromptSelector

selector = QLearningPromptSelector(["What is AI?", "Explain machine learning"])
next_prompt = selector.select_next_prompt("What is AI?", passed_guard=True)
selector.update_rewards("What is AI?", next_prompt, reward=1.0, passed_guard=True)