MAINT: Store hash in prompt template identifiers

pyrit/converter/prompt_template_converter.py

@@ -1,6 +1,7 @@
 # Copyright (c) Microsoft Corporation.
 # Licensed under the MIT license.
 
+import hashlib
 import re
 
 from pyrit.converter.converter import Converter, ConverterResult
@@ -16,6 +17,10 @@ class PromptTemplateConverter(Converter):
     framing (e.g. ``TASK is '{{ prompt }}'``) as well as wrapping the prompt in
     surrounding content, such as hiding it in an HTML comment for indirect prompt
     injection (e.g. ``<p>Visible text</p><!-- {{ prompt }} -->``).
+
+    The converter identifier stores a fixed-size hash of the configured template,
+    rather than the template text. The deprecated ``TaskFramingConverter`` retains
+    its legacy identifier.
     """
 
     SUPPORTED_INPUT_TYPES = ("text",)
@@ -56,9 +61,10 @@ def _build_identifier(self) -> ComponentIdentifier:
         Returns:
             ComponentIdentifier: The identifier for this converter.
         """
+        template_hash = hashlib.sha256(self._template.encode("utf-8")).hexdigest()[:16]
         return self._create_identifier(
             params={
-                "template": self._template,
+                "template_hash": template_hash,
                 "strip_characters": self._strip_characters,
             },
         )

tests/unit/converter/test_prompt_template_converter.py

@@ -96,12 +96,24 @@ def test_input_output_types():
     assert converter.output_supported("text") is True
 
 
-def test_identifier_includes_template_and_strip_characters():
+def test_identifier_includes_template_hash_and_strip_characters():
     converter = PromptTemplateConverter(template="T: {{ prompt }}", strip_characters="'")
     identifier = converter.get_identifier()
     assert identifier.class_name == "PromptTemplateConverter"
-    assert identifier.params["template"] == "T: {{ prompt }}"
+    assert identifier.params["template_hash"] == "1d6417505e0571d2"
     assert identifier.params["strip_characters"] == "'"
+    assert "template" not in identifier.params
+
+
+def test_identifier_keeps_hash():
+    converter = PromptTemplateConverter(template="T: {{ prompt }}", strip_characters="'")
+    assert converter.get_identifier().hash == "f7a37b712954fc0dd196ee9481ca7a1e0779841b768171bb6af146a046dde693"
+
+
+def test_identifier_matches_for_equivalent_templates():
+    first = PromptTemplateConverter(template="A {{ prompt }}").get_identifier()
+    second = PromptTemplateConverter(template="A {{ prompt }}").get_identifier()
+    assert first == second
 
 
 def test_identifier_differs_by_template():
@@ -110,6 +122,19 @@ def test_identifier_differs_by_template():
     assert first.hash != second.hash
 
 
+def test_identifier_differs_by_strip_characters():
+    first = PromptTemplateConverter(template="A {{ prompt }}", strip_characters="'").get_identifier()
+    second = PromptTemplateConverter(template="A {{ prompt }}", strip_characters='"').get_identifier()
+    assert first.hash != second.hash
+
+
+def test_long_template_identifier_uses_fixed_size_hash():
+    template = f"{'A' * 10_000} {{{{ prompt }}}}"
+    identifier = PromptTemplateConverter(template=template).get_identifier()
+    assert identifier.params["template_hash"] == "a280c02c55a808af"
+    assert template not in identifier.model_dump_json()
+
+
 # Markdown hidden-text recipes from the converters doc: put the prompt on one line,
 # backslash-escape ``\`` and the delimiter, then insert it with PromptTemplateConverter.
 _MARKDOWN_RECIPES = {

tests/unit/converter/test_task_framing_converter.py

@@ -56,3 +56,10 @@ def test_identifier_keeps_task_template_param_name():
     assert identifier.params["task_template"] == "TASK is '{{ prompt }}'"
     assert identifier.params["strip_characters"] == "'"
     assert "template" not in identifier.params
+    assert "template_hash" not in identifier.params
+
+
+def test_identifier_keeps_legacy_hash():
+    # Pinned to the v1.1.0 value so the deprecated converter keeps matching identifiers already stored.
+    identifier = _make(strip_characters="'").get_identifier()
+    assert identifier.hash == "b680cca8b216deec74b66758dd6bdc005fe21e593fbd0a9a975fe40e7e9fe56d"