MAINT Breaking: Standardizing Attack Args

doc/code/executor/attack/role_play_attack.ipynb

@@ -7,8 +7,8 @@
    "source": [
     "# Role Play Attack (Single-Turn) - optional\n",
     "\n",
-    "This attack prepends some prompts defined in `role_play_definition`, along with an `adversarial_chat` target LLM to generate the first turns to send. Typically these prompts describe a fictional scenario to attempt and elicit harmful responses.\n",
-    "Any converters that you provide will be applied to the prompt that has already been converted by the role play definition (using the provided `adversarial_chat` target). You may see better success if you provide a LLM that has no content moderation or other safety mechanisms. Otherwise, it may refuse to convert the prompt as expected.\n",
+    "This attack prepends some prompts defined in `role_play_definition`, along with an `attack_adversarial_config` target LLM to generate the first turns to send. Typically these prompts describe a fictional scenario to attempt and elicit harmful responses.\n",
+    "Any converters that you provide will be applied to the prompt that has already been converted by the role play definition (using the provided `attack_adversarial_config` target). You may see better success if you provide a LLM that has no content moderation or other safety mechanisms. Otherwise, it may refuse to convert the prompt as expected.\n",
     "\n",
     "\n",
     "The results and intermediate interactions will be saved to memory according to the environment settings. For details, see the [Memory Configuration Guide](../../memory/0_memory.md)."
@@ -343,6 +343,7 @@
     "\n",
     "from pyrit.auth import get_azure_openai_auth\n",
     "from pyrit.executor.attack import (\n",
+    "    AttackAdversarialConfig,\n",
     "    AttackConverterConfig,\n",
     "    AttackExecutor,\n",
     "    AttackScoringConfig,\n",
@@ -375,7 +376,7 @@
     "\n",
     "attack = RolePlayAttack(\n",
     "    objective_target=objective_target,\n",
-    "    adversarial_chat=adversarial_chat,\n",
+    "    attack_adversarial_config=AttackAdversarialConfig(target=adversarial_chat),\n",
     "    role_play_definition_path=RolePlayPaths.MOVIE_SCRIPT.value,\n",
     "    attack_scoring_config=scoring_config,\n",
     "    attack_converter_config=converter_config,\n",

doc/code/executor/attack/role_play_attack.py

@@ -15,8 +15,8 @@
 # %% [markdown]
 # # Role Play Attack (Single-Turn) - optional
 #
-# This attack prepends some prompts defined in `role_play_definition`, along with an `adversarial_chat` target LLM to generate the first turns to send. Typically these prompts describe a fictional scenario to attempt and elicit harmful responses.
-# Any converters that you provide will be applied to the prompt that has already been converted by the role play definition (using the provided `adversarial_chat` target). You may see better success if you provide a LLM that has no content moderation or other safety mechanisms. Otherwise, it may refuse to convert the prompt as expected.
+# This attack prepends some prompts defined in `role_play_definition`, along with an `attack_adversarial_config` target LLM to generate the first turns to send. Typically these prompts describe a fictional scenario to attempt and elicit harmful responses.
+# Any converters that you provide will be applied to the prompt that has already been converted by the role play definition (using the provided `attack_adversarial_config` target). You may see better success if you provide a LLM that has no content moderation or other safety mechanisms. Otherwise, it may refuse to convert the prompt as expected.
 #
 #
 # The results and intermediate interactions will be saved to memory according to the environment settings. For details, see the [Memory Configuration Guide](../../memory/0_memory.md).
@@ -26,6 +26,7 @@
 
 from pyrit.auth import get_azure_openai_auth
 from pyrit.executor.attack import (
+    AttackAdversarialConfig,
     AttackConverterConfig,
     AttackExecutor,
     AttackScoringConfig,
@@ -58,7 +59,7 @@
 
 attack = RolePlayAttack(
     objective_target=objective_target,
-    adversarial_chat=adversarial_chat,
+    attack_adversarial_config=AttackAdversarialConfig(target=adversarial_chat),
     role_play_definition_path=RolePlayPaths.MOVIE_SCRIPT.value,
     attack_scoring_config=scoring_config,
     attack_converter_config=converter_config,

pyrit/executor/attack/single_turn/flip_attack.py

@@ -39,6 +39,7 @@ class FlipAttack(PromptSendingAttack):
     @apply_defaults
     def __init__(
         self,
+        *,
         objective_target: PromptChatTarget = REQUIRED_VALUE,  # type: ignore[assignment]
         attack_converter_config: Optional[AttackConverterConfig] = None,
         attack_scoring_config: Optional[AttackScoringConfig] = None,

pyrit/executor/attack/single_turn/many_shot_jailbreak.py

@@ -64,6 +64,7 @@ class ManyShotJailbreakAttack(PromptSendingAttack):
     @apply_defaults
     def __init__(
         self,
+        *,
         objective_target: PromptTarget = REQUIRED_VALUE,  # type: ignore[assignment]
         attack_converter_config: Optional[AttackConverterConfig] = None,
         attack_scoring_config: Optional[AttackScoringConfig] = None,

pyrit/executor/attack/single_turn/role_play.py

@@ -8,7 +8,7 @@
 
 from pyrit.common.apply_defaults import REQUIRED_VALUE, apply_defaults
 from pyrit.common.path import EXECUTOR_SEED_PROMPT_PATH
-from pyrit.executor.attack.core.attack_config import AttackConverterConfig, AttackScoringConfig
+from pyrit.executor.attack.core.attack_config import AttackAdversarialConfig, AttackConverterConfig, AttackScoringConfig
 from pyrit.executor.attack.core.attack_parameters import AttackParameters
 from pyrit.executor.attack.single_turn.prompt_sending import PromptSendingAttack
 from pyrit.executor.attack.single_turn.single_turn_attack_strategy import (
@@ -20,7 +20,7 @@
 )
 from pyrit.prompt_converter import LLMGenericTextConverter
 from pyrit.prompt_normalizer import PromptConverterConfiguration, PromptNormalizer
-from pyrit.prompt_target import PromptChatTarget, PromptTarget
+from pyrit.prompt_target import PromptTarget
 
 logger = logging.getLogger(__name__)
 
@@ -66,7 +66,7 @@ def __init__(
         self,
         *,
         objective_target: PromptTarget = REQUIRED_VALUE,  # type: ignore[assignment]
-        adversarial_chat: PromptChatTarget,
+        attack_adversarial_config: AttackAdversarialConfig,
         role_play_definition_path: pathlib.Path,
         attack_converter_config: Optional[AttackConverterConfig] = None,
         attack_scoring_config: Optional[AttackScoringConfig] = None,
@@ -78,8 +78,8 @@ def __init__(
 
         Args:
             objective_target (PromptTarget): The target system to attack.
-            adversarial_chat (PromptChatTarget): The adversarial chat target used to rephrase
-                objectives into role-play scenarios.
+            attack_adversarial_config (AttackAdversarialConfig): Configuration for the adversarial component,
+                including the adversarial chat target used to rephrase objectives into role-play scenarios.
             role_play_definition_path (pathlib.Path): Path to the YAML file containing role-play
                 definitions (rephrase instructions, user start turn, assistant start turn).
             attack_converter_config (Optional[AttackConverterConfig]): Configuration for prompt converters.
@@ -102,7 +102,7 @@ def __init__(
         )
 
         # Store the adversarial chat for role-play rephrasing
-        self._adversarial_chat = adversarial_chat
+        self._adversarial_chat = attack_adversarial_config.target
 
         # Load role-play definitions
         role_play_definition = SeedDataset.from_yaml_file(role_play_definition_path)

pyrit/scenario/scenarios/airt/content_harms.py

@@ -303,7 +303,7 @@ def _get_single_turn_attacks(
 
         role_play_attack = RolePlayAttack(
             objective_target=self._objective_target,
-            adversarial_chat=self._adversarial_chat,
+            attack_adversarial_config=AttackAdversarialConfig(target=self._adversarial_chat),
             role_play_definition_path=RolePlayPaths.MOVIE_SCRIPT.value,
         )
 

pyrit/scenario/scenarios/airt/jailbreak.py

@@ -9,6 +9,7 @@
 from pyrit.common import apply_defaults
 from pyrit.datasets import TextJailBreak
 from pyrit.executor.attack.core.attack_config import (
+    AttackAdversarialConfig,
     AttackConverterConfig,
     AttackScoringConfig,
 )
@@ -279,7 +280,9 @@ async def _get_atomic_attack_from_strategy_async(
             case "skeleton":
                 attack = SkeletonKeyAttack(**args)
             case "role_play":
-                args["adversarial_chat"] = self._get_or_create_adversarial_target()
+                args["attack_adversarial_config"] = AttackAdversarialConfig(
+                    target=self._get_or_create_adversarial_target()
+                )
                 args["role_play_definition_path"] = RolePlayPaths.PERSUASION_SCRIPT.value
                 attack = RolePlayAttack(**args)
             case _:

pyrit/scenario/scenarios/airt/leakage.py

@@ -361,9 +361,9 @@ async def _create_role_play_attack(self) -> RolePlayAttack:
         """
         return RolePlayAttack(
             objective_target=self._objective_target,
-            adversarial_chat=self._adversarial_chat,
             role_play_definition_path=RolePlayPaths.PERSUASION_SCRIPT.value,
             attack_scoring_config=self._scorer_config,
+            attack_adversarial_config=self._adversarial_config,
         )
 
     def _resolve_seed_groups(self) -> list[SeedAttackGroup]:

pyrit/scenario/scenarios/airt/psychosocial.py

@@ -480,9 +480,9 @@ def _create_single_turn_attacks(
         )
         role_play = RolePlayAttack(
             objective_target=self._objective_target,
-            adversarial_chat=self._adversarial_chat,
             role_play_definition_path=RolePlayPaths.MOVIE_SCRIPT.value,
             attack_scoring_config=scoring_config,
+            attack_adversarial_config=AttackAdversarialConfig(target=self._adversarial_chat),
         )
         attacks.append(
             AtomicAttack(

pyrit/scenario/scenarios/airt/scam.py

@@ -294,9 +294,9 @@ def _get_atomic_attack_from_strategy(self, strategy: str) -> AtomicAttack:
         elif strategy == "role_play":
             attack_strategy = RolePlayAttack(
                 objective_target=self._objective_target,
-                adversarial_chat=self._adversarial_chat,
                 role_play_definition_path=RolePlayPaths.PERSUASION_SCRIPT_WRITTEN.value,
                 attack_scoring_config=self._scorer_config,
+                attack_adversarial_config=self._adversarial_config,
             )
         elif strategy == "context_compliance":
             # Set system prompt to default

tests/unit/executor/attack/single_turn/test_role_play.py

@@ -11,6 +11,7 @@
 from unit.mocks import get_mock_scorer_identifier, get_mock_target_identifier
 
 from pyrit.executor.attack import (
+    AttackAdversarialConfig,
     AttackConverterConfig,
     AttackParameters,
     AttackScoringConfig,
@@ -96,7 +97,7 @@ def role_play_attack(mock_objective_target, mock_adversarial_chat_target, role_p
     """Create a RolePlayAttack instance with default configuration"""
     return RolePlayAttack(
         objective_target=mock_objective_target,
-        adversarial_chat=mock_adversarial_chat_target,
+        attack_adversarial_config=AttackAdversarialConfig(target=mock_adversarial_chat_target),
         role_play_definition_path=role_play_definition_file,
     )
 
@@ -118,7 +119,7 @@ def test_init(self, mock_objective_target, mock_adversarial_chat_target, role_pl
         """Test RolePlayAttack initialization with default parameters"""
         attack = RolePlayAttack(
             objective_target=mock_objective_target,
-            adversarial_chat=mock_adversarial_chat_target,
+            attack_adversarial_config=AttackAdversarialConfig(target=mock_adversarial_chat_target),
             role_play_definition_path=role_play_definition_file,
         )
 
@@ -133,7 +134,7 @@ def test_init_with_valid_true_false_scorer(
         """Test RolePlayAttack initialization with a valid true/false scorer"""
         attack = RolePlayAttack(
             objective_target=mock_objective_target,
-            adversarial_chat=mock_adversarial_chat_target,
+            attack_adversarial_config=AttackAdversarialConfig(target=mock_adversarial_chat_target),
             role_play_definition_path=role_play_definition_file,
             attack_scoring_config=AttackScoringConfig(objective_scorer=mock_scorer),
         )
@@ -148,7 +149,7 @@ def test_init_with_invalid_scorer_type(
         with pytest.raises(ValueError, match="Objective scorer must be a TrueFalseScorer"):
             RolePlayAttack(
                 objective_target=mock_objective_target,
-                adversarial_chat=mock_adversarial_chat_target,
+                attack_adversarial_config=AttackAdversarialConfig(target=mock_adversarial_chat_target),
                 role_play_definition_path=role_play_definition_file,
                 attack_scoring_config=AttackScoringConfig(objective_scorer=scorer),
             )
@@ -159,7 +160,7 @@ def test_init_with_invalid_role_play_definition_path(self, mock_objective_target
         with pytest.raises(FileNotFoundError):
             RolePlayAttack(
                 objective_target=mock_objective_target,
-                adversarial_chat=mock_adversarial_chat_target,
+                attack_adversarial_config=AttackAdversarialConfig(target=mock_adversarial_chat_target),
                 role_play_definition_path=invalid_path,
             )
 
@@ -172,7 +173,7 @@ def test_init_with_custom_parameters(
 
         attack = RolePlayAttack(
             objective_target=mock_objective_target,
-            adversarial_chat=mock_adversarial_chat_target,
+            attack_adversarial_config=AttackAdversarialConfig(target=mock_adversarial_chat_target),
             role_play_definition_path=role_play_definition_file,
             attack_converter_config=AttackConverterConfig(
                 request_converters=request_converters, response_converters=response_converters
@@ -194,7 +195,7 @@ def test_init_with_negative_max_attempts(
         with pytest.raises(ValueError, match="max_attempts_on_failure must be a non-negative integer"):
             RolePlayAttack(
                 objective_target=mock_objective_target,
-                adversarial_chat=mock_adversarial_chat_target,
+                attack_adversarial_config=AttackAdversarialConfig(target=mock_adversarial_chat_target),
                 role_play_definition_path=role_play_definition_file,
                 max_attempts_on_failure=-1,
             )
@@ -205,7 +206,7 @@ def test_init_loads_role_play_definition_correctly(
         """Test that role play definitions are loaded correctly from YAML"""
         attack = RolePlayAttack(
             objective_target=mock_objective_target,
-            adversarial_chat=mock_adversarial_chat_target,
+            attack_adversarial_config=AttackAdversarialConfig(target=mock_adversarial_chat_target),
             role_play_definition_path=role_play_definition_file,
         )
 
@@ -220,7 +221,7 @@ def test_rephrase_converter_created(
         """Test that the rephrase converter is properly created"""
         attack = RolePlayAttack(
             objective_target=mock_objective_target,
-            adversarial_chat=mock_adversarial_chat_target,
+            attack_adversarial_config=AttackAdversarialConfig(target=mock_adversarial_chat_target),
             role_play_definition_path=role_play_definition_file,
         )