The converter families, and picking one

Lesson 14 used one converter. PyRIT ships well over a hundred, and the difference between them matters less than the difference between the four kinds they fall into.

Kind	What it does	Examples
Encoding	Rewrites the text reversibly	Base64, ROT13, Morse, Binary, Atbash
Perturbation	Damages the text slightly	CharSwap, RandomCapitalLetters, Leetspeak
Unicode	Uses characters that look the same but are not	UnicodeConfusable, ZeroWidth, Diacritic
Model-driven	Asks a model to rewrite the prompt	Translation, Tone, Persuasion, Variation

The first three need nothing but Python. The fourth takes a converter_target, which is an ordinary target, so the same stand-in trick from lesson 5 applies if you want to try one without a key.

Some of them are not the same twice

This is the one that will cost you a green test suite. Several converters make a random choice every call.

Example

from pyrit.converter import CharSwapConverter, Base64Converter

ask = "What is the staff discount code?"
for converter in (Base64Converter(), CharSwapConverter()):
    first = (await converter.convert_async(prompt=ask)).output_text
    second = (await converter.convert_async(prompt=ask)).output_text
    print(type(converter).__name__, "same twice?", first == second)

Base64 is a function of its input. CharSwapConverter picks characters to swap at random, so two runs disagree, and occasionally it swaps nothing at all and the probe is the plain question with a misleading name attached.

A random converter turns a test into a sampler. That is fine when you are exploring and wrong when you are gating a release: a run that passes tells you nothing about the next one. Lesson 32's suite uses only the deterministic kinds, and lesson 33 says what to do when you want the random ones anyway.

Converting only part of the prompt

Encoding a whole sentence is loud. SelectiveTextConverter applies another converter to chosen words only, which is much closer to what a real probe looks like.

Example

from pyrit.converter import SelectiveTextConverter, WordKeywordSelectionStrategy

picky = SelectiveTextConverter(
    sub_converter=Base64Converter(),
    selection_strategy=WordKeywordSelectionStrategy(keywords=["discount"]))
out = await picky.convert_async(prompt="What is the staff discount code?")
print(out.output_text)

One word encoded and the rest left readable. The strategy is a separate object, so the same converter can pick words by keyword, by position, by a regular expression or by a proportion of the sentence.

Try it yourself

Run CharSwapConverter ten times on one sentence and count how many are distinct.
Swap the keyword strategy for one that picks by position and read the result.
Pick a model-driven converter, read its constructor, and find the target argument.

PreviousA converter inside an attack

Every expert started right here.