diff --git a/pyrit/prompt_target/prompt_shield_target.py b/pyrit/prompt_target/prompt_shield_target.py
index 0c9cf411c3..6dbf7d617e 100644
--- a/pyrit/prompt_target/prompt_shield_target.py
+++ b/pyrit/prompt_target/prompt_shield_target.py
@@ -311,23 +311,31 @@ def _input_parser(self, input_str: str) -> dict[str, Any]:
if TYPE_CHECKING:
assert_never(unknown_field)
- # This can also be accomplished using regex, but Python is flexible.
- user_prompt = ""
+ # Scan linearly for matched ... pairs so unmatched closing tags
+ # cannot misclassify userPrompt text as documents or drop trailing prompt text.
+ open_tag = ""
+ close_tag = ""
+ user_prompt_parts: list[str] = []
documents: list[str] = []
- split_input = input_str.split("")
+ cursor = 0
- for element in split_input:
- contents = element.split("")
+ while True:
+ open_idx = input_str.find(open_tag, cursor)
+ if open_idx == -1:
+ user_prompt_parts.append(input_str[cursor:])
+ break
- if len(contents) == 1:
- user_prompt += contents[0]
- else:
- # Text before the closing tag is the document; any text after it
- # belongs to the user prompt and must not be silently dropped.
- documents.append(contents[0])
- user_prompt += contents[1]
+ content_start = open_idx + len(open_tag)
+ close_idx = input_str.find(close_tag, content_start)
+ if close_idx == -1:
+ user_prompt_parts.append(input_str[cursor:])
+ break
+
+ user_prompt_parts.append(input_str[cursor:open_idx])
+ documents.append(input_str[content_start:close_idx])
+ cursor = close_idx + len(close_tag)
- return {"userPrompt": user_prompt, "documents": documents if documents else []}
+ return {"userPrompt": "".join(user_prompt_parts), "documents": documents}
def _add_auth_param_to_headers(self, headers: dict[str, str]) -> None:
"""
diff --git a/tests/unit/prompt_target/target/test_prompt_shield_target.py b/tests/unit/prompt_target/target/test_prompt_shield_target.py
index aaa038e3f0..49ca236891 100644
--- a/tests/unit/prompt_target/target/test_prompt_shield_target.py
+++ b/tests/unit/prompt_target/target/test_prompt_shield_target.py
@@ -106,6 +106,53 @@ async def test_prompt_shield_document_parsing_keeps_text_between_documents(promp
assert result["documents"] == [" d1 ", " d2 "]
+@pytest.mark.parametrize(
+ ("input_str", "expected"),
+ [
+ (
+ "hello world",
+ {"userPrompt": "hello world", "documents": []},
+ ),
+ (
+ "doc1 mid tail",
+ {"userPrompt": " mid tail", "documents": ["doc1"]},
+ ),
+ (
+ "a b c",
+ {"userPrompt": "a b c", "documents": []},
+ ),
+ (
+ "Explain the tag",
+ {"userPrompt": "Explain the tag", "documents": []},
+ ),
+ (
+ "Summarize: Here is a tag please",
+ {"userPrompt": "Summarize: please", "documents": ["Here is a tag"]},
+ ),
+ (
+ "",
+ {"userPrompt": "", "documents": [""]},
+ ),
+ ],
+ ids=[
+ "stray_closing_tag",
+ "extra_closing_tag_after_document",
+ "multiple_stray_closing_tags",
+ "unclosed_opening_tag",
+ "literal_opening_tag_inside_document",
+ "empty_document",
+ ],
+)
+def test_prompt_shield_document_parsing_handles_unmatched_and_nested_tags(
+ promptshield_target: PromptShieldTarget,
+ input_str: str,
+ expected: dict[str, object],
+) -> None:
+ result = promptshield_target._input_parser(input_str)
+
+ assert result == expected
+
+
def test_prompt_shield_unknown_entry_field_preserves_runtime_fallback(promptshield_target: PromptShieldTarget):
object.__setattr__(promptshield_target, "_force_entry_field", "future-field")