You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
{{ message }}
Repository navigation
DOC: Document XPIAManualProcessingWorkflow on the Workflows page #4499
The Workflows page documents XPIAWorkflow and XPIATestWorkflow for cross-domain prompt injection (XPIA) testing. A third variant, XPIAManualProcessingWorkflow, is exported from pyrit.executor.workflow but is not mentioned anywhere in the docs.
It is the variant for when the system that processes the injected content cannot be driven from code: the workflow plants the attack, pauses, asks the operator to trigger the processing step by hand (for example by opening the document in the application under test), and then takes the output pasted into the console and scores it.
What to change
Add a section to doc/code/executor/5_workflow.py that covers:
When to choose it over XPIAWorkflow and XPIATestWorkflow.
What the operator sees and has to do at the pause.
A code example showing how to construct and run it (attack_setup_target and scorer are required).
The workflow waits for console input, so it cannot run when the notebook is executed. Show the example as a fenced code block inside a markdown cell, not as an executed cell.
Done when
The page describes all three XPIA workflow variants and when to use each.
The example matches the constructor's current signature.
The .py and .ipynb files match, and pre-commit passes.
Read the class and its unit tests first, and describe only what the code does today.
Keep the text short and concrete. Lead with when someone would use the class, then show the smallest example that works.
pre-commit runs the doc validation hooks; they must pass.
How to contribute
Claim it first. Before you start, comment on this issue to say you are taking it.
First comment wins. If yours is the first claim comment, go ahead and open a PR; you do not need to wait for a reply. If someone has already commented, this issue is taken, so please pick a different one.
I'll take this one. Since #4499 has no comments yet, I'll claim it here and open a PR.
Quick plan: add a short section to doc/code/executor/5_workflow.py after the existing XPIA content, framing it as when-to-use (auto-driven target vs. test target vs. human-in-the-loop), with the manual example in a fenced code block inside a markdown cell so the notebook doesn't block on input(). Then regenerate the matching .ipynb via jupytext and run pre-commit so both stay in sync. I'll read xpia.py and its unit tests first and only describe current behavior, and link this issue in the PR.
One thing to confirm: is pasting the operator output back into the console still the supported path for XPIAManualProcessingWorkflow, or should the docs steer people to a non-interactive scoring hook if one exists?
Activity
galiliya commentedon Nov 21, 2026
I'll take this one. Since #4499 has no comments yet, I'll claim it here and open a PR.
Quick plan: add a short section to
doc/code/executor/5_workflow.pyafter the existing XPIA content, framing it as when-to-use (auto-driven target vs. test target vs. human-in-the-loop), with the manual example in a fenced code block inside a markdown cell so the notebook doesn't block oninput(). Then regenerate the matching.ipynbvia jupytext and run pre-commit so both stay in sync. I'll readxpia.pyand its unit tests first and only describe current behavior, and link this issue in the PR.One thing to confirm: is pasting the operator output back into the console still the supported path for
XPIAManualProcessingWorkflow, or should the docs steer people to a non-interactive scoring hook if one exists?