[Shivam Sharma (ShivamSharma43)]

Description

Fixes #3478.

PromptShieldTarget._send_prompt_to_target_async parsed and sent request.original_value, the prompt from before any request converter ran. Every other target sends converted_value, so a converter applied through PromptNormalizer or an attack was silently ignored: testing whether Prompt Shield catches a Base64 or leetspeak variant of an injection sent the plain original prompt instead. The <document> split was also done on the unconverted text.

#3100 fixed the same problem in PromptShieldScorer. This is the matching change in the target.

Changes:

• PromptShieldTarget passes request.converted_value to _input_parser, so both userPrompt and documents come from the converted text.
• The docstring says the converted text is parsed.

A piece with no converter has converted_value == original_value, so unconverted prompts are sent exactly as before. PromptShieldScorer is unaffected: it already puts the converted text into the piece it sends.

Tests and Documentation

• New tests in tests/unit/prompt_target/target/test_prompt_shield_target.py:
  • send_prompt_async with a piece whose converted_value differs from original_value sends the parsed converted value, including its <document> part. A piece without a converted value still sends the original.
  • PromptNormalizer.send_prompt_async with a Base64Converter sends the Base64 text to Prompt Shield.
• On main the two converted-value cases fail (the original prompt is sent); with this change they pass. The unconverted case passes on both.
• pytest tests/unit/prompt_target tests/unit/score: 6380 passed, 369 skipped.
• ruff check, ruff format and ty check pass on the changed files.
• No documentation changes beyond the docstring.

🤖 Generated with Claude Code

[Copilot]

Copilot was unable to review this pull request because the user who requested the review has reached their quota limit.