Skip to content

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

.

A-Vamshi0f35e1bb2adb
diff --git a/deepteam/confident/api.py b/deepteam/confident/api.py
index 18709d92..42f37d12 100644
--- a/deepteam/confident/api.py
+++ b/deepteam/confident/api.py
@@ -24,6 +24,7 @@ from deepeval.confident.api import (
class Endpoints(Enum):
RISK_ASSESSMENT_ENDPOINT = "/v1/risk-assessments"
+ RT_FRAMEWORK_ENDPOINT = "/v1/risk-assessments/frameworks/:frameworkId"
class Api:
diff --git a/deepteam/red_team.py b/deepteam/red_team.py
index 81e7383e..a901d8bc 100644
--- a/deepteam/red_team.py
+++ b/deepteam/red_team.py
@@ -6,14 +6,14 @@ from deepteam.attacks import BaseAttack
from deepteam.attacks.multi_turn.types import CallbackType
from deepteam.red_teamer import RedTeamer
from deepteam.attacks.attack_engine import AttackEngine
-from deepteam.frameworks.frameworks import AISafetyFramework
+from deepteam.frameworks.frameworks import RedTeamingFramework
def red_team(
model_callback: CallbackType,
vulnerabilities: Optional[List[BaseVulnerability]] = None,
attacks: Optional[List[BaseAttack]] = None,
- framework: Optional[AISafetyFramework] = None,
+ framework: Optional[RedTeamingFramework] = None,
simulator_model: DeepEvalBaseLLM = "gpt-4o-mini",
evaluation_model: DeepEvalBaseLLM = "gpt-4o-mini",
attacks_per_vulnerability_type: int = 1,
diff --git a/deepteam/red_teamer/api.py b/deepteam/red_teamer/api.py
index 413c8f43..0f16b703 100644
--- a/deepteam/red_teamer/api.py
+++ b/deepteam/red_teamer/api.py
@@ -61,6 +61,7 @@ class APIRiskAssessment(BaseModel):
run_duration: float = Field(alias="runDuration")
identifier: Optional[str] = Field(alias="identifier")
assessment_cost: Optional[float] = Field(alias="assessmentCost")
+ rt_framework_id: Optional[str] = Field(None, alias="rtFrameworkId")
test_cases: List[APIRTTestCase] = Field(alias="testCases")
@@ -140,6 +141,7 @@ def map_risk_assessment_to_api(
risk_assessment: RiskAssessment,
assessment_cost: Optional[float] = None,
identifier: Optional[str] = None,
+ rt_framework_id: Optional[str] = None,
) -> APIRiskAssessment:
"""
Map RiskAssessment to APIRiskAssessment.
@@ -148,6 +150,8 @@ def map_risk_assessment_to_api(
risk_assessment: The internal RiskAssessment object
identifier: Optional[str] = None,
assessment_cost: Optional[float] = None,
+ rt_framework_id: id of the Confident AI framework this assessment ran,
+ set only when the framework was pulled from Confident AI.
Returns:
APIRiskAssessment: The API-compatible risk assessment object
@@ -173,5 +177,6 @@ def map_risk_assessment_to_api(
runDuration=risk_assessment.overview.run_duration,
identifier=identifier,
assessmentCost=assessment_cost,
+ rtFrameworkId=rt_framework_id,
testCases=test_cases,
)
diff --git a/deepteam/red_teamer/red_teamer.py b/deepteam/red_teamer/red_teamer.py
index ca6ed2d3..4b439c34 100644
--- a/deepteam/red_teamer/red_teamer.py
+++ b/deepteam/red_teamer/red_teamer.py
@@ -18,7 +18,7 @@ from deepteam.red_teamer.api import map_risk_assessment_to_api
from deepteam.test_case import RTTestCase, RTTurn
from deepeval.utils import get_or_create_event_loop
-from deepteam.frameworks.frameworks import AISafetyFramework
+from deepteam.frameworks.frameworks import RedTeamingFramework
from deepteam.frameworks.risk_category import RiskCategory
from deepteam.telemetry import capture_red_teamer_run
from deepteam.attacks import BaseAttack
@@ -97,7 +97,7 @@ class RedTeamer:
attacks: Optional[List[BaseAttack]] = None,
simulator_model: DeepEvalBaseLLM = None,
evaluation_model: DeepEvalBaseLLM = None,
- framework: Optional[AISafetyFramework] = None,
+ framework: Optional[RedTeamingFramework] = None,
attacks_per_vulnerability_type: int = 1,
ignore_errors: bool = True,
reuse_simulated_test_cases: bool = False,
@@ -165,7 +165,7 @@ class RedTeamer:
)
if _upload_to_confident:
self.risk_assessment = risk_assessment
- self._post_risk_assessment()
+ self._post_risk_assessment(rt_framework_id=framework._id)
return risk_assessment
if framework and framework._has_dataset:
@@ -302,7 +302,9 @@ class RedTeamer:
if _print_assessment:
self._print_risk_assessment(self.risk_assessment)
if _upload_to_confident:
- self._post_risk_assessment()
+ self._post_risk_assessment(
+ rt_framework_id=framework._id if framework else None
+ )
return self.risk_assessment
@@ -313,7 +315,7 @@ class RedTeamer:
attacks: Optional[List[BaseAttack]] = None,
simulator_model: DeepEvalBaseLLM = None,
evaluation_model: DeepEvalBaseLLM = None,
- framework: Optional[AISafetyFramework] = None,
+ framework: Optional[RedTeamingFramework] = None,
attacks_per_vulnerability_type: int = 1,
ignore_errors: bool = False,
reuse_simulated_test_cases: bool = False,
@@ -359,7 +361,7 @@ class RedTeamer:
)
if _upload_to_confident:
self.risk_assessment = risk_assessment
- self._post_risk_assessment()
+ self._post_risk_assessment(rt_framework_id=framework._id)
return risk_assessment
if framework:
@@ -507,7 +509,9 @@ class RedTeamer:
if _print_assessment:
self._print_risk_assessment(self.risk_assessment)
if _upload_to_confident:
- self._post_risk_assessment()
+ self._post_risk_assessment(
+ rt_framework_id=framework._id if framework else None
+ )
return self.risk_assessment
@@ -863,7 +867,7 @@ class RedTeamer:
console.print("[bold magenta]LLM red teaming complete.[/bold magenta]")
console.print("=" * 80 + "\n")
- def _post_risk_assessment(self):
+ def _post_risk_assessment(self, rt_framework_id: Optional[str] = None):
if not is_confident():
passing = 0
failing = 0
@@ -887,7 +891,9 @@ class RedTeamer:
return
api = Api()
- api_risk_assessment = map_risk_assessment_to_api(self.risk_assessment)
+ api_risk_assessment = map_risk_assessment_to_api(
+ self.risk_assessment, rt_framework_id=rt_framework_id
+ )
try:
body = api_risk_assessment.model_dump(
by_alias=True, exclude_none=True
@@ -1026,7 +1032,7 @@ class RedTeamer:
model_callback: CallbackType,
simulator_model: DeepEvalBaseLLM = None,
evaluation_model: DeepEvalBaseLLM = None,
- framework: Optional[AISafetyFramework] = None,
+ framework: Optional[RedTeamingFramework] = None,
attacks_per_vulnerability_type: int = 1,
ignore_errors: bool = False,
reuse_simulated_test_cases: bool = False,
@@ -1069,6 +1075,10 @@ class RedTeamer:
total=1,
)
framework_assessment = assess_risk_category(risk_category)
+ # The framework's own risk category is more specific than
+ # the one `red_team` derives from each vulnerability type.
+ for test_case in framework_assessment.test_cases:
+ test_case.risk_category = risk_category.name
results[risk_category.name] = framework_assessment
update_pbar(progress_2, risk_task_id, advance_to_end=True)
update_pbar(progress, task_id)
@@ -1095,7 +1105,7 @@ class RedTeamer:
model_callback: CallbackType,
simulator_model: DeepEvalBaseLLM = None,
evaluation_model: DeepEvalBaseLLM = None,
- framework: Optional[AISafetyFramework] = None,
+ framework: Optional[RedTeamingFramework] = None,
attacks_per_vulnerability_type: int = 1,
ignore_errors: bool = False,
reuse_simulated_test_cases: bool = False,
@@ -1130,6 +1140,10 @@ class RedTeamer:
_print_assessment=False,
_upload_to_confident=False,
)
+ # The framework's own risk category is more specific than
+ # the one `a_red_team` derives from each vulnerability type.
+ for test_case in assessment.test_cases:
+ test_case.risk_category = category.name
update_pbar(progress_2, risk_task_id, advance_to_end=True)
return category.name, assessment
diff --git a/docs/content/docs/red-teaming-introduction.mdx b/docs/content/docs/red-teaming-introduction.mdx
index 5212b824..d9951ba1 100644
--- a/docs/content/docs/red-teaming-introduction.mdx
+++ b/docs/content/docs/red-teaming-introduction.mdx
@@ -440,7 +440,7 @@ There is **ONE** mandatory and **ELEVEN** optional arguments when calling the `r
- `model_callback`: a callback of type `Callable[[str], str]` that wraps around the target LLM system you wish to red team.
- [Optional] `vulnerabilities`: a list of type `BaseVulnerability`s that determines the weaknesses to detect for.
- [Optional] `attacks`: a list of type `BaseAttack`s that determines the methods that will be simulated to expose the defined `vulnerabilities`.
-- [Optional] `framework`: an object of type `AISafetyFramework` that contains both `vulnerabilities` and `attacks` in it.
+- [Optional] `framework`: an object of type `RedTeamingFramework` that contains both `vulnerabilities` and `attacks` in it.
- [Optional] `simulator_model`: a string specifying which of OpenAI's GPT models to use, **OR** [any custom LLM model](https://docs.confident-ai.com/guides/guides-using-custom-llms) of type `DeepEvalBaseLLM` for simulating attacks. Defaulted to `"gpt-3.5-turbo-0125"`.
- [Optional] `evaluation_model`: a string specifying which of OpenAI's GPT models to use, **OR** [any custom LLM model](https://docs.confident-ai.com/guides/guides-using-custom-llms) of type `DeepEvalBaseLLM` for evaluation. Defaulted to `"gpt-4o"`.
- [Optional] `attacks_per_vulnerability_type`: an int that determines the number of attacks to be simulated per vulnerability type. Defaulted to `1`.
@@ -593,7 +593,7 @@ As explained in the adversarial attack section, by making the PromptInjection at
There is **ONE** mandatory and **TEN** optional arguments when calling the `red_team()` method:
- `model_callback`: a callback of type `Callable[[str], str]` that wraps around the target LLM system you wish to red team.
-- [Optional] `framework`: an object of type `AISafetyFramework` that contains both `vulnerabilities` and `attacks` in it.
+- [Optional] `framework`: an object of type `RedTeamingFramework` that contains both `vulnerabilities` and `attacks` in it.
- [Optional] `simulator_model`: a string specifying which of OpenAI's GPT models to use, **OR** [any custom LLM model](https://docs.confident-ai.com/guides/guides-using-custom-llms) of type `DeepEvalBaseLLM` for simulating attacks. Defaulted to `"gpt-3.5-turbo-0125"`.
- [Optional] `evaluation_model`: a string specifying which of OpenAI's GPT models to use, **OR** [any custom LLM model](https://docs.confident-ai.com/guides/guides-using-custom-llms) of type `DeepEvalBaseLLM` for evaluation. Defaulted to `"gpt-4o"`.
- [Optional] `attacks_per_vulnerability_type`: an int that determines the number of attacks to be simulated per vulnerability type. Defaulted to `1`.