Skip to content

DeepTeam is a framework to red team LLMs and AI agents.

update single-turn constants to pass parameters

A-Vamshic01f588ea4ad
diff --git a/deepteam/attacks/single_turn/__init__.py b/deepteam/attacks/single_turn/__init__.py
index fc83810b..15c97b8f 100644
--- a/deepteam/attacks/single_turn/__init__.py
+++ b/deepteam/attacks/single_turn/__init__.py
@@ -1,4 +1,4 @@
-from .base_single_turn_attack import BaseSingleTurnAttack
+from .base_single_turn_attack import BaseSingleTurnAttack, AttackParameter
from .adversarial_poetry import AdversarialPoetry
from .base64 import Base64
from .gray_box import GrayBox
@@ -26,6 +26,7 @@ from .context_poisoning.context_poisoning import ContextPoisoning
__all__ = [
"AdversarialPoetry",
+ "AttackParameter",
"BaseSingleTurnAttack",
"Base64",
"GrayBox",
diff --git a/deepteam/attacks/single_turn/authority_escalation/authority_escalation.py b/deepteam/attacks/single_turn/authority_escalation/authority_escalation.py
index 8bd25a8a..a5633c85 100644
--- a/deepteam/attacks/single_turn/authority_escalation/authority_escalation.py
+++ b/deepteam/attacks/single_turn/authority_escalation/authority_escalation.py
@@ -4,7 +4,7 @@ from deepeval.models import DeepEvalBaseLLM
from deepeval.metrics.utils import initialize_model
from deepteam.utils import create_progress, update_pbar, add_pbar
-from deepteam.attacks.single_turn import BaseSingleTurnAttack
+from deepteam.attacks.single_turn import BaseSingleTurnAttack, AttackParameter
from deepteam.attacks.base_attack import Exploitability
from deepteam.attacks.single_turn.authority_escalation.template import (
AuthorityEscalationTemplate,
@@ -27,6 +27,12 @@ class AuthorityEscalation(BaseSingleTurnAttack):
name = "Authority Escalation"
exploitability = Exploitability.HIGH
description = "Rewrites the attack to mimic a superior, administrator, or compliance officer, using authoritative language to bypass restrictions."
+ parameters = {
+ "role": AttackParameter(
+ type="string",
+ description="Authority role to emulate during rewrite (for example, admin or compliance officer).",
+ )
+ }
def __init__(
self,
diff --git a/deepteam/attacks/single_turn/base_single_turn_attack.py b/deepteam/attacks/single_turn/base_single_turn_attack.py
index 4e1df8bf..6ef68e91 100644
--- a/deepteam/attacks/single_turn/base_single_turn_attack.py
+++ b/deepteam/attacks/single_turn/base_single_turn_attack.py
@@ -1,5 +1,20 @@
+from typing import Any, Dict, List, Literal, Optional
+
+from pydantic import BaseModel
+
from deepteam.attacks.base_attack import BaseAttack
+ParameterType = Literal["string", "integer", "float", "boolean", "enum", "json"]
+
+
+class AttackParameter(BaseModel):
+ type: ParameterType
+ required: bool = False
+ default: Optional[Any] = None
+ options: Optional[List[str]] = None
+ description: Optional[str] = None
+
+
class BaseSingleTurnAttack(BaseAttack):
- pass
+ parameters: Dict[str, AttackParameter] = {}
diff --git a/deepteam/attacks/single_turn/constants.py b/deepteam/attacks/single_turn/constants.py
index 108a663d..d68029f0 100644
--- a/deepteam/attacks/single_turn/constants.py
+++ b/deepteam/attacks/single_turn/constants.py
@@ -1,7 +1,7 @@
from typing import Dict
-from pydantic import BaseModel
+from pydantic import BaseModel, Field
-from .base_single_turn_attack import BaseSingleTurnAttack
+from .base_single_turn_attack import BaseSingleTurnAttack, AttackParameter
from .base64 import Base64
from .gray_box import GrayBox
from .leetspeak import Leetspeak
@@ -62,6 +62,7 @@ SINGLE_TURN_ATTACK_NAMES = sorted(
class AttackInfo(BaseModel):
description: str
exploitability: str
+ parameters: Dict[str, AttackParameter] = Field(default_factory=dict)
# Map attack names to their description and exploitability
@@ -69,6 +70,7 @@ SINGLE_TURN_ATTACK_INFO_MAP: Dict[str, AttackInfo] = {
name: AttackInfo(
description=attack_class.description,
exploitability=attack_class.exploitability.value,
+ parameters=attack_class.parameters,
)
for name, attack_class in SINGLE_TURN_ATTACK_CLASSES_MAP.items()
}
diff --git a/deepteam/attacks/single_turn/context_flooding/context_flooding.py b/deepteam/attacks/single_turn/context_flooding/context_flooding.py
index 24a4e068..611cd25d 100644
--- a/deepteam/attacks/single_turn/context_flooding/context_flooding.py
+++ b/deepteam/attacks/single_turn/context_flooding/context_flooding.py
@@ -1,5 +1,5 @@
from typing import Literal, Optional
-from deepteam.attacks.single_turn import BaseSingleTurnAttack
+from deepteam.attacks.single_turn import BaseSingleTurnAttack, AttackParameter
from deepteam.attacks.base_attack import Exploitability
from deepteam.attacks.single_turn.context_flooding.template import (
DEFAULT_PREFIXES,
@@ -12,6 +12,23 @@ class ContextFlooding(BaseSingleTurnAttack):
name = "ContextFlooding"
exploitability = Exploitability.MEDIUM
description = "A single-turn attack that floods the LLM's context with long pre-defined prefixes to hit a target character size."
+ parameters = {
+ "target_chars": AttackParameter(
+ type="integer",
+ default=4000,
+ description="Target character length for generated context prefix.",
+ ),
+ "prefix": AttackParameter(
+ type="string",
+ description="Custom prefix text; when set, overrides prefix_type.",
+ ),
+ "prefix_type": AttackParameter(
+ type="enum",
+ default="enterprise",
+ options=["enterprise", "logs", "transcript"],
+ description="Built-in prefix template to use when custom prefix is not provided.",
+ ),
+ }
def __init__(
self,
diff --git a/deepteam/attacks/single_turn/embedded_instruction_json/embedded_instruction_json.py b/deepteam/attacks/single_turn/embedded_instruction_json/embedded_instruction_json.py
index cc6c07b7..c36651b5 100644
--- a/deepteam/attacks/single_turn/embedded_instruction_json/embedded_instruction_json.py
+++ b/deepteam/attacks/single_turn/embedded_instruction_json/embedded_instruction_json.py
@@ -2,7 +2,7 @@ import json
import copy
from typing import Optional, Dict, Any, Literal
-from deepteam.attacks.single_turn import BaseSingleTurnAttack
+from deepteam.attacks.single_turn import BaseSingleTurnAttack, AttackParameter
from deepteam.attacks.base_attack import Exploitability
from deepteam.attacks.single_turn.embedded_instruction_json.template import (
DEFAULT_JSON,
@@ -15,6 +15,28 @@ class EmbeddedInstructionJSON(BaseSingleTurnAttack):
name = "EmbeddedInstructionJSON"
exploitability = Exploitability.MEDIUM
description = "A single-turn attack that embeds an attack payload as structured JSON data."
+ parameters = {
+ "target_chars": AttackParameter(
+ type="integer",
+ default=2000,
+ description="Target character size for generated JSON context.",
+ ),
+ "target_depth": AttackParameter(
+ type="integer",
+ default=4,
+ description="Depth of nested structure used for embedding payload.",
+ ),
+ "json_template": AttackParameter(
+ type="json",
+ description="Optional base JSON template used before expansion.",
+ ),
+ "attack_position": AttackParameter(
+ type="enum",
+ default="end",
+ options=["start", "middle", "end"],
+ description="Where deep_context payload is inserted into the JSON object.",
+ ),
+ }
def __init__(
self,
diff --git a/deepteam/attacks/single_turn/emotional_manipulation/emotional_manipulation.py b/deepteam/attacks/single_turn/emotional_manipulation/emotional_manipulation.py
index 6865f70b..9a0952db 100644
--- a/deepteam/attacks/single_turn/emotional_manipulation/emotional_manipulation.py
+++ b/deepteam/attacks/single_turn/emotional_manipulation/emotional_manipulation.py
@@ -4,7 +4,7 @@ from deepeval.models import DeepEvalBaseLLM
from deepeval.metrics.utils import initialize_model
from deepteam.utils import create_progress, update_pbar, add_pbar
-from deepteam.attacks.single_turn import BaseSingleTurnAttack
+from deepteam.attacks.single_turn import BaseSingleTurnAttack, AttackParameter
from deepteam.attacks.base_attack import Exploitability
from deepteam.attacks.single_turn.emotional_manipulation.template import (
EmotionalManipulationTemplate,
@@ -29,6 +29,19 @@ class EmotionalManipulation(BaseSingleTurnAttack):
name = "Emotional Manipulation"
exploitability = Exploitability.HIGH
description = "Rewrites the attack by wrapping it in a high-intensity emotional narrative (Anger, Pity, Urgency) to bypass safety filters."
+ parameters = {
+ "emotion": AttackParameter(
+ type="enum",
+ options=[
+ "ANGER",
+ "PITY",
+ "CONFUSION",
+ "AUTHORITY",
+ "URGENCY",
+ ],
+ description="Emotional strategy applied while rewriting the attack.",
+ )
+ }
def __init__(
self,
diff --git a/deepteam/attacks/single_turn/multilingual/multilingual.py b/deepteam/attacks/single_turn/multilingual/multilingual.py
index d7a4d93d..692a7baa 100644
--- a/deepteam/attacks/single_turn/multilingual/multilingual.py
+++ b/deepteam/attacks/single_turn/multilingual/multilingual.py
@@ -4,7 +4,7 @@ from deepeval.models import DeepEvalBaseLLM
from deepeval.metrics.utils import initialize_model
from deepteam.utils import create_progress, update_pbar, add_pbar
-from deepteam.attacks.single_turn import BaseSingleTurnAttack
+from deepteam.attacks.single_turn import BaseSingleTurnAttack, AttackParameter
from deepteam.attacks.base_attack import Exploitability
from deepteam.attacks.single_turn.multilingual.template import (
MultilingualTemplate,
@@ -27,6 +27,12 @@ class Multilingual(BaseSingleTurnAttack):
name = "Multilingual"
exploitability = Exploitability.MEDIUM
description = "A translation-based attack that converts prompts into low-resource or non-English languages to exploit weaker safety training in multilingual models."
+ parameters = {
+ "language": AttackParameter(
+ type="string",
+ description="Language used for the intermediate rewrite before returning English output.",
+ )
+ }
def __init__(
self,
diff --git a/deepteam/attacks/single_turn/roleplay/roleplay.py b/deepteam/attacks/single_turn/roleplay/roleplay.py
index 7aa2772f..079268f1 100644
--- a/deepteam/attacks/single_turn/roleplay/roleplay.py
+++ b/deepteam/attacks/single_turn/roleplay/roleplay.py
@@ -4,7 +4,7 @@ from typing import Optional, Union
from deepeval.metrics.utils import initialize_model
from deepeval.models import DeepEvalBaseLLM
-from deepteam.attacks.single_turn import BaseSingleTurnAttack
+from deepteam.attacks.single_turn import BaseSingleTurnAttack, AttackParameter
from deepteam.attacks.base_attack import Exploitability
from deepteam.attacks.single_turn.roleplay.template import (
RoleplayTemplate,
@@ -28,6 +28,16 @@ class Roleplay(BaseSingleTurnAttack):
name = "Roleplay"
exploitability = Exploitability.MEDIUM
description = "A persona-based attack that instructs the model to adopt a fictional character, expert role, or alternate identity to justify generating harmful content."
+ parameters = {
+ "persona": AttackParameter(
+ type="string",
+ description="Fictional persona to adopt while rewriting the attack.",
+ ),
+ "role": AttackParameter(
+ type="string",
+ description="Role framing used to justify the persona context.",
+ ),
+ }
def __init__(
self,
diff --git a/deepteam/attacks/single_turn/synthetic_context_injection/synthetic_context_injection.py b/deepteam/attacks/single_turn/synthetic_context_injection/synthetic_context_injection.py
index 23da780a..1c91d2a3 100644
--- a/deepteam/attacks/single_turn/synthetic_context_injection/synthetic_context_injection.py
+++ b/deepteam/attacks/single_turn/synthetic_context_injection/synthetic_context_injection.py
@@ -4,7 +4,7 @@ from deepeval.models import DeepEvalBaseLLM
from deepeval.metrics.utils import initialize_model
from deepteam.utils import create_progress, update_pbar, add_pbar
-from deepteam.attacks.single_turn import BaseSingleTurnAttack
+from deepteam.attacks.single_turn import BaseSingleTurnAttack, AttackParameter
from deepteam.attacks.base_attack import Exploitability
from deepteam.attacks.single_turn.synthetic_context_injection.template import (
SyntheticContextInjectionTemplate,
@@ -27,6 +27,17 @@ class SyntheticContextInjection(BaseSingleTurnAttack):
name = "Synthetic Context Injection"
exploitability = Exploitability.HIGH
description = "Generates domain-specific 'fake history' that mimics internal company logs, emails, and verification states."
+ parameters = {
+ "target_information": AttackParameter(
+ type="string",
+ required=True,
+ description="Target information the synthetic context should center around.",
+ ),
+ "context_style": AttackParameter(
+ type="string",
+ description="Optional style or domain framing for generated context.",
+ ),
+ }
def __init__(
self,