Skip to content

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

add types

A-Vamshidf4f59de82e2
diff --git a/deepteam/red_teamer/api.py b/deepteam/red_teamer/api.py
index 413c8f43..8c560af1 100644
--- a/deepteam/red_teamer/api.py
+++ b/deepteam/red_teamer/api.py
@@ -29,6 +29,8 @@ class AttackMethodResult(BaseModel):
class APIRTTurn(TurnApi):
turn_level_attack: Optional[str] = Field(None, alias="turnLevelAttack")
+ stopping_reason: Optional[str] = Field(None, alias="stoppingReason")
+ stopping_category: Optional[str] = Field(None, alias="stoppingCategory")
class APIRTTestCase(BaseModel):
@@ -70,10 +72,12 @@ def map_turn_to_api(turn: RTTurn, order: int) -> APIRTTurn:
role=turn.role,
content=turn.content,
order=order,
- user_id=turn.user_id,
- retrieval_context=turn.retrieval_context,
- tools_called=turn.tools_called,
- turn_level_attack=turn.turn_level_attack,
+ userId=turn.user_id,
+ retrievalContext=turn.retrieval_context,
+ toolsCalled=turn.tools_called,
+ turnLevelAttack=turn.turn_level_attack,
+ stoppingReason=turn.stopping_reason,
+ stoppingCategory=turn.stopping_category,
)
diff --git a/deepteam/red_teamer/red_teamer.py b/deepteam/red_teamer/red_teamer.py
index ca6ed2d3..493f1176 100644
--- a/deepteam/red_teamer/red_teamer.py
+++ b/deepteam/red_teamer/red_teamer.py
@@ -36,6 +36,8 @@ from deepteam.vulnerabilities import BaseVulnerability
from deepteam.vulnerabilities.types import VulnerabilityType
from deepteam.attacks.attack_simulator import AttackSimulator
from deepteam.attacks.attack_engine import AttackEngine
+from deepteam.attacks.attack_simulator.utils import add_cost
+from deepteam.attacks.multi_turn.progression import StopReason
from deepteam.attacks.multi_turn.types import CallbackType
from deepteam.metrics import BaseRedTeamingMetric
from deepteam.red_teamer.risk_assessment import (
@@ -47,6 +49,14 @@ from deepteam.risks import getRiskCategory
console = Console()
+def _has_detected_shift(test_case: RTTestCase) -> bool:
+ if test_case.score is None or not test_case.turns:
+ return False
+ return (
+ test_case.turns[-1].stopping_category == StopReason.SHIFT_DETECTED.value
+ )
+
+
class RedTeamer:
risk_assessment: Optional[RiskAssessment] = None
simulated_test_cases: Optional[List[RTTestCase]] = None
@@ -205,6 +215,9 @@ class RedTeamer:
simulated_test_cases = framework.test_cases
else:
self.attack_simulator.model_callback = model_callback
+ self.attack_simulator.evaluation_model = (
+ self.evaluation_model
+ )
if vulnerabilities:
self.attack_simulator.attack_engine = (
self._resolve_attack_engine(attack_engine)
@@ -401,6 +414,9 @@ class RedTeamer:
simulated_test_cases = framework.test_cases
else:
self.attack_simulator.model_callback = model_callback
+ self.attack_simulator.evaluation_model = (
+ self.evaluation_model
+ )
if vulnerabilities:
self.attack_simulator.attack_engine = (
self._resolve_attack_engine(attack_engine)
@@ -539,11 +555,17 @@ class RedTeamer:
if simulated_test_case.error is not None:
return red_teaming_test_case
+ if _has_detected_shift(red_teaming_test_case):
+ return red_teaming_test_case
+
try:
metric.measure(red_teaming_test_case)
red_teaming_test_case.score = metric.score
red_teaming_test_case.reason = metric.reason
- red_teaming_test_case.evaluation_cost = metric.evaluation_cost
+ red_teaming_test_case.evaluation_cost = add_cost(
+ red_teaming_test_case.evaluation_cost,
+ metric.evaluation_cost,
+ )
except:
if ignore_errors:
red_teaming_test_case.error = f"Error evaluating target LLM output for the '{vulnerability_type.value}' vulnerability type"
@@ -622,11 +644,17 @@ class RedTeamer:
if red_teaming_test_case.error is not None:
return red_teaming_test_case
+ if _has_detected_shift(red_teaming_test_case):
+ return red_teaming_test_case
+
try:
await metric.a_measure(red_teaming_test_case)
red_teaming_test_case.score = metric.score
red_teaming_test_case.reason = metric.reason
- red_teaming_test_case.evaluation_cost = metric.evaluation_cost
+ red_teaming_test_case.evaluation_cost = add_cost(
+ red_teaming_test_case.evaluation_cost,
+ metric.evaluation_cost,
+ )
except:
if ignore_errors:
red_teaming_test_case.error = f"Error evaluating target LLM output for the '{vulnerability_type.value}' vulnerability type"
diff --git a/deepteam/test_case/test_case.py b/deepteam/test_case/test_case.py
index 7d939723..919cb1b9 100644
--- a/deepteam/test_case/test_case.py
+++ b/deepteam/test_case/test_case.py
@@ -1,4 +1,3 @@
-from nntplib import NNTPDataError
from typing import List, Optional
from enum import Enum
from deepeval.test_case import LLMTestCase, Turn, ToolCall
@@ -6,11 +5,17 @@ from deepeval.test_case import LLMTestCase, Turn, ToolCall
class RTTurn(Turn):
turn_level_attack: Optional[str] = None
+ stopping_reason: Optional[str] = None
+ stopping_category: Optional[str] = None
def __repr__(self):
attrs = [f"role={self.role}", f"content={self.content}"]
if self.turn_level_attack is not None:
attrs.append(f"turn_level_attack={self.turn_level_attack}")
+ if self.stopping_category is not None:
+ attrs.append(f"stopping_category={self.stopping_category}")
+ if self.stopping_reason is not None:
+ attrs.append(f"stopping_reason={self.stopping_reason}")
return f"RTTurn({', '.join(attrs)})"