Bug: PromptNormalizer.send_prompt_async returns None on skip criteria, crashing downstream callers

[pr0b3r7]

Bug Report

Describe the bug

PromptNormalizer.send_prompt_async() returns None when skip criteria match (lines 89-90 of prompt_normalizer.py), but callers assume a non-None PromptRequestResponse is always returned. This causes AttributeError: 'NoneType' object has no attribute 'get_value' in downstream code like LLMGenericTextConverter.convert_async().

Steps to reproduce

1. Create a PromptNormalizer instance
2. Call set_skip_criteria() with criteria that match a prompt
3. Call send_prompt_async() — returns None
4. Any caller that accesses response.get_value() or response.request_pieces crashes

Root cause analysis

File: pyrit/prompt_normalizer/prompt_normalizer.py

# Lines 89-90: Returns None when skip criteria match
if self._should_skip_based_on_skip_criteria(request):
    return None  # <-- Callers assume non-None return

# Lines 124-125: Defensive None return (unreachable in practice)
if response is None:
    return None

Crash site — File: pyrit/prompt_converter/llm_generic_text_converter.py

# Lines 99-100: No None check before accessing .get_value()
response = await self._converter_target.send_prompt_async(prompt_request=request)
return ConverterResult(output_text=response.get_value(), output_type="text")
# ^^^^ AttributeError when response is None

Partial mitigation already exists: send_prompt_batch_to_target_async() at line 185 correctly filters out None returns:

return [response for response in responses if response is not None]

But single-prompt callers (converters, orchestrators) have no such protection.

Downstream impact

When azure-ai-evaluation's OrchestratorManager uses PyRIT converters that hit this path:

1. The AttributeError propagates up through PyRIT's exception handler
2. Gets wrapped as "Error sending prompt with conversation ID: ..." (line 122)
3. The Azure SDK's retry decorator misclassifies this as a network error
4. Retries 5 times with exponential backoff (~47 seconds wasted, all retries fail identically)

A companion issue has been filed on Azure/azure-sdk-for-python for the retry-misclassification side.

Proposed fix

Option A (Preferred): Return a sentinel empty PromptRequestResponse instead of None:

if self._should_skip_based_on_skip_criteria(request):
    skipped = construct_response_from_request(
        request=request.request_pieces[0],
        response_text_pieces=[""],
        response_type="text",
        error="skipped",
    )
    return skipped

Option B: Change return type to Optional[PromptRequestResponse] and update all callers to handle None.

Option C: Raise a specific PromptSkippedException that callers can catch.

Environment

• PyRIT version: 0.8.1 (also verified present in 0.11.0 via GitHub source)
• Python: 3.12.12
• OS: macOS 15.5 (Darwin 24.6.0)
• azure-ai-evaluation: 1.15.0

Additional context

Discovered during red teaming framework development when using LLMGenericTextConverter with skip criteria enabled via azure-ai-evaluation's RedTeam class. The bug is silent in batch operations (filtered by list comprehension) but crashes single-prompt converter paths.