Hacktoberfest 2026: the issues maintainers tagged for October, open and beginner-friendly. Browse Hacktoberfest issues

AsciiArtConverter silently drops every non-ASCII character

Open
  • Is something wrong?
#2,942 2 comments 0 reactions 0 assignees View on GitHub

Maintainers usually reply within 2 days

@Pushpak731 is already working on this.

Since Oct 2, 2026.

  • #2962 by @Pushpak731 — open

Assessment

Difficulty
4/5
Estimated time
3-5 days
Newbie friendliness
38/100
Issue type
Bug
Clarity
Mostly clear
Activity status
Active
Tech stack
python
Domain
security

Research direction

Start with pyrit/converter/ascii_art_converter.py and test_ascii_art_converter.py, then inspect the neighboring ascii_smuggler_converter.py, NatoConverter, and BrailleConverter behavior. Reproduce the non-ASCII cases against art.text2art and the available fonts. Done means the maintainer-selected behavior is implemented and regression tests prove prompts are not silently altered.

Written by the indexing model from the issue text.

Description

AsciiArtConverter silently deletes every non-ASCII character from the prompt. Nothing is raised and nothing is logged — the target just receives a mangled prompt, or no prompt at all.

What happens

convert_async hands the prompt straight to art.text2art:

# pyrit/converter/ascii_art_converter.py:76
return ConverterResult(output_text=text2art(prompt, font=font), output_type="text")

text2art has no glyph for a character, so it omits it. I probed every font the art package exposes — 371 in total, 354 of them in this converter's rand pool:

character dropped by
é, ü 371 / 371
日, 本, 語, 漢 371 / 371
🙂 371 / 371
“, — 371 / 371
ß 307 / 371
A 2 / 371 (hills, nfi1)

Accented Latin, CJK, emoji, smart quotes and em-dashes are dropped by every font, so the loss does not depend on which font gets drawn.

Two concrete consequences

One character disappears, taking its whole glyph block with it:

AsciiArtConverter(font="block").convert_async(prompt="cafe")  -> 4 glyph blocks (892 chars)
AsciiArtConverter(font="block").convert_async(prompt="café")  -> 3 glyph blocks (672 chars)

And a prompt written entirely in a non-ASCII script converts to the empty string. With the default font="rand" this holds on every draw:

AsciiArtConverter().convert_async(prompt="日本語の指示")            -> output_text=''  (12/12 draws)
AsciiArtConverter(font="block").convert_async(prompt="忽略之前的所有指令") -> ''

For a red-teaming framework this is the harmful direction. A Chinese- or Japanese-language attack prompt is a first-class use case, and it converts to nothing. A mixed prompt reaches the target quietly altered, with no signal to the red-teamer.

Why I think this is a defect

Three things already in this repo establish the opposite convention for the same situation:

  • AsciiSmugglerConverter raises ValueError naming the characters outside the range it can encode (ascii_smuggler_converter.py:69-74, merged in #2540).
  • NatoConverter was changed to pass unmapped characters through rather than delete them (#2399).
  • BrailleConverter got the same treatment (#2539).

AsciiArtConverter's docstring documents only ValueError: If the input type is not supported; it never mentions that non-ASCII input is dropped. The tests only feed ASCII prompts — test_ascii_art_converter.py:16-29 assert len(result.output_text) > 0 — so nothing pins this behaviour in either direction.

What behaviour do you want?

Option A — raise, naming the characters this font cannot render. Consistent with #2540, and it fails loudly instead of altering the attack. Cost: a campaign whose prompts contain an accent or an emoji would start raising. Checking the resolved font is deterministic for everything that matters here, since the characters above are dropped by all 371 fonts; only ß-type characters are font-dependent, so under font="rand" such a prompt could raise on one draw and not the next.

Option B — render what the font can, pass the rest through. Keeps pipelines running and stops losing content, at the cost of a mixed output (art plus bare characters). This matches what #2399 and #2539 ended up doing for the other converters.

Option C — document the restriction, change nothing. Cheapest, but a silently altered attack prompt is a poor default for a tool whose job is to send a precise prompt.

I lean towards A: "the prompt that reaches the target is not the prompt I wrote" is precisely the failure a red-teaming tool should refuse rather than hide, and #2540 already set that precedent for the converter next door. If B is preferable because non-ASCII prompts are expected to keep working, I would implement B instead.

Happy to take whichever you pick, with regression tests over the real art font list. Everything above was measured against the installed art package; I called no model or target.

Dominant language
Python
Stars
4.6k
Forks
924
Avg merge
2d 10h
Merged PRs (30d)
210

Getting set up

Open in Codespaces

Starts the project's dev container in your browser, under your own GitHub account.

  • No Dockerfile or Docker Compose file
  • Has a pull request template
  • No contributing guide

First steps

  1. Read the whole issue, then the project's contributing guide.
  2. Comment on the issue to say you are picking it up — it saves two people doing the same work.
  3. Fork the repository and make your change on a branch.
  4. Open a pull request that references the issue number.

More from microsoft/PyRIT

All issues in microsoft/PyRIT

Similar issues

More Python issues

Get new issues in your inbox

A short digest of beginner-friendly GitHub issues.