Skip to content

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

removed the metrics map

A-Vamshidd90afdf4054
diff --git a/deepteam/red_teamer/red_teamer.py b/deepteam/red_teamer/red_teamer.py
index fe57947a..db9fa914 100644
--- a/deepteam/red_teamer/red_teamer.py
+++ b/deepteam/red_teamer/red_teamer.py
@@ -172,8 +172,6 @@ class RedTeamer:
vulnerabilities=[v.get_name() for v in vulnerabilities],
attacks=[a.get_name() for a in attacks],
):
- # Initialize metric map
- metrics_map = self.get_red_teaming_metrics_map(vulnerabilities)
# Simulate attacks
if (
reuse_simulated_attacks
@@ -208,9 +206,12 @@ class RedTeamer:
)
for vulnerability_type, simulated_attacks in pbar:
for simulated_attack in simulated_attacks:
- metric: BaseRedTeamingMetric = metrics_map.get(
- vulnerability_type
- )()
+ for vulnerability in vulnerabilities:
+ if vulnerability_type in vulnerability.types:
+ metric: BaseRedTeamingMetric = vulnerability._get_metric(
+ vulnerability_type
+ )
+ break
red_teaming_test_case = RedTeamingTestCase(
vulnerability=simulated_attack.vulnerability,
@@ -301,9 +302,6 @@ class RedTeamer:
vulnerabilities=[v.get_name() for v in vulnerabilities],
attacks=[a.get_name() for a in attacks],
):
- # Initialize metric map
- metrics_map = self.get_red_teaming_metrics_map(vulnerabilities)
-
# Generate attacks
if (
reuse_simulated_attacks
@@ -362,9 +360,9 @@ class RedTeamer:
async with semaphore:
test_cases = await self._a_evaluate_vulnerability_type(
model_callback,
+ vulnerabilities,
vulnerability_type,
attacks,
- metrics_map,
ignore_errors=ignore_errors,
)
red_teaming_test_cases.extend(test_cases)
@@ -398,7 +396,7 @@ class RedTeamer:
simulated_attack: SimulatedAttack,
vulnerability: str,
vulnerability_type: VulnerabilityType,
- metrics_map,
+ vulnerabilities: List[BaseVulnerability],
ignore_errors: bool,
) -> RedTeamingTestCase:
red_teaming_test_case = RedTeamingTestCase(
@@ -413,7 +411,12 @@ class RedTeamer:
red_teaming_test_case.error = simulated_attack.error
return red_teaming_test_case
- metric: BaseRedTeamingMetric = metrics_map[vulnerability_type]()
+ for _vulnerability in vulnerabilities:
+ if vulnerability_type in _vulnerability.types:
+ metric: BaseRedTeamingMetric = _vulnerability._get_metric(
+ vulnerability_type
+ )
+ break
try:
actual_output = await model_callback(simulated_attack.input)
red_teaming_test_case.actual_output = actual_output
@@ -445,9 +448,9 @@ class RedTeamer:
async def _a_evaluate_vulnerability_type(
self,
model_callback: CallbackType,
+ vulnerabilities: List[BaseVulnerability],
vulnerability_type: VulnerabilityType,
simulated_attacks: List[SimulatedAttack],
- metrics_map,
ignore_errors: bool,
) -> List[RedTeamingTestCase]:
red_teaming_test_cases = await asyncio.gather(
@@ -455,9 +458,9 @@ class RedTeamer:
self._a_attack(
model_callback=model_callback,
simulated_attack=simulated_attack,
+ vulnerabilities=vulnerabilities,
vulnerability=simulated_attack.vulnerability,
vulnerability_type=vulnerability_type,
- metrics_map=metrics_map,
ignore_errors=ignore_errors,
)
for simulated_attack in simulated_attacks
@@ -465,223 +468,6 @@ class RedTeamer:
)
return red_teaming_test_cases
- ##################################################
- ### Metrics Map ##################################
- ##################################################
-
- def get_red_teaming_metrics_map(
- self, vulnerabilities: List[BaseVulnerability]
- ):
-
- metrics_map = {
- #### Bias ####
- **{
- bias_type: lambda: BiasMetric(
- model=self.evaluation_model,
- purpose=self.target_purpose,
- async_mode=self.async_mode,
- )
- for bias_type in BiasType
- },
- #### Toxicity ####
- **{
- toxicity_type: lambda tt=toxicity_type: ToxicityMetric(
- model=self.evaluation_model,
- toxicity_category=tt.value,
- async_mode=self.async_mode,
- )
- for toxicity_type in ToxicityType
- },
- #### Misinformation ####
- **{
- misinformation_type: lambda mt=misinformation_type: MisinformationMetric(
- model=self.evaluation_model,
- misinformation_category=mt.value,
- async_mode=self.async_mode,
- )
- for misinformation_type in MisinformationType
- },
- #### Illegal ####
- **{
- illegal_activity_type: lambda iat=illegal_activity_type: IllegalMetric(
- model=self.evaluation_model,
- illegal_category=iat.value,
- async_mode=self.async_mode,
- )
- for illegal_activity_type in IllegalActivityType
- },
- #### Prompt Leakage ####
- **{
- prompt_leakage_type: lambda: PromptExtractionMetric(
- model=self.evaluation_model,
- purpose=self.target_purpose,
- async_mode=self.async_mode,
- )
- for prompt_leakage_type in PromptLeakageType
- },
- #### PII Leakage ####
- **{
- pii_type: lambda: PIIMetric(
- model=self.evaluation_model,
- purpose=self.target_purpose,
- async_mode=self.async_mode,
- )
- for pii_type in PIILeakageType
- },
- #### Debug Access ####
- **{
- debug_access_type: lambda: DebugAccessMetric(
- model=self.evaluation_model, async_mode=self.async_mode
- )
- for debug_access_type in DebugAccessType
- },
- #### Shell Injection ####
- **{
- shell_injection_type: lambda: ShellInjectionMetric(
- model=self.evaluation_model, async_mode=self.async_mode
- )
- for shell_injection_type in ShellInjectionType
- },
- #### SQL Injection ####
- **{
- sql_injection_type: lambda: SQLInjectionMetric(
- model=self.evaluation_model, async_mode=self.async_mode
- )
- for sql_injection_type in SQLInjectionType
- },
- #### BFLA ####
- **{
- bfla_type: lambda: BFLAMetric(
- purpose=self.target_purpose,
- model=self.evaluation_model,
- async_mode=self.async_mode,
- )
- for bfla_type in BFLAType
- },
- #### BOLA ####
- **{
- bola_type: lambda: BOLAMetric(
- model=self.evaluation_model,
- async_mode=self.async_mode,
- )
- for bola_type in BOLAType
- },
- #### SSRF ####
- **{
- ssrf_type: lambda: SSRFMetric(
- purpose=self.target_purpose,
- model=self.evaluation_model,
- async_mode=self.async_mode,
- )
- for ssrf_type in SSRFType
- },
- #### RBAC ####
- **{
- rbac_type: lambda: RBACMetric(
- model=self.evaluation_model,
- purpose=self.target_purpose,
- async_mode=self.async_mode,
- )
- for rbac_type in RBACType
- },
- #### Excessive Agency ####
- **{
- excessive_agency_type: lambda: ExcessiveAgencyMetric(
- model=self.evaluation_model,
- purpose=self.target_purpose,
- async_mode=self.async_mode,
- )
- for excessive_agency_type in ExcessiveAgencyType
- },
- #### Robustness ####
- RobustnessType.HIJACKING: lambda: HijackingMetric(
- purpose=self.target_purpose,
- model=self.evaluation_model,
- async_mode=self.async_mode,
- ),
- RobustnessType.INPUT_OVERRELIANCE: lambda: OverrelianceMetric(
- purpose=self.target_purpose,
- model=self.evaluation_model,
- async_mode=self.async_mode,
- ),
- #### Intellectual Property ####
- **{
- ip_type: lambda: IntellectualPropertyMetric(
- model=self.evaluation_model,
- purpose=self.target_purpose,
- async_mode=self.async_mode,
- )
- for ip_type in IntellectualPropertyType
- },
- #### Competition ####
- **{
- competiton_type: lambda: CompetitorsMetric(
- model=self.evaluation_model,
- purpose=self.target_purpose,
- async_mode=self.async_mode,
- )
- for competiton_type in CompetitionType
- },
- #### Graphic Content ####
- **{
- content_type: lambda ct=content_type: GraphicMetric(
- model=self.evaluation_model,
- graphic_category=ct.value,
- async_mode=self.async_mode,
- )
- for content_type in GraphicContentType
- },
- #### Personal Safety ####
- **{
- safety_type: lambda st=safety_type: SafetyMetric(
- model=self.evaluation_model,
- safety_category=st.value,
- async_mode=self.async_mode,
- )
- for safety_type in PersonalSafetyType
- },
- #### Agentic Vulnerabilities ####
- **{
- subversion_type: lambda: SubversionSuccessMetric(
- model=self.evaluation_model,
- purpose=self.target_purpose,
- async_mode=self.async_mode,
- )
- for subversion_type in RecursiveHijackingType
- },
- **{
- extraction_type: lambda: ExtractionSuccessMetric(
- model=self.evaluation_model,
- purpose=self.target_purpose,
- async_mode=self.async_mode,
- )
- for extraction_type in GoalTheftType
- },
- }
-
- # Store custom vulnerability instances for dynamic metric assignment
- for vulnerability in vulnerabilities:
- if isinstance(vulnerability, CustomVulnerability):
- for vuln_type in vulnerability.get_types():
- metric = vulnerability.get_metric()
- if metric:
- metrics_map[vuln_type] = lambda: metric
- else:
- criteria = vulnerability.get_criteria()
- if not criteria:
- raise ValueError(
- f"CustomVulnerability '{vulnerability.get_name()}' must provide a 'criteria' parameter that defines what should be evaluated."
- )
-
- metrics_map[vuln_type] = lambda hc=criteria: HarmMetric(
- model=self.evaluation_model,
- harm_category=hc,
- async_mode=self.async_mode,
- )
-
- self.metrics_map = metrics_map
- return metrics_map
-
def save_test_cases_as_simulated_attacks(
self, test_cases: List[RedTeamingTestCase]
):
diff --git a/deepteam/vulnerabilities/agentic/excessive_agency/excessive_agency.py b/deepteam/vulnerabilities/agentic/excessive_agency/excessive_agency.py
index afced966..d51acb80 100644
--- a/deepteam/vulnerabilities/agentic/excessive_agency/excessive_agency.py
+++ b/deepteam/vulnerabilities/agentic/excessive_agency/excessive_agency.py
@@ -14,6 +14,7 @@ ExcessiveAgencyLiteral = Literal["functionality", "permissions", "autonomy"]
class ExcessiveAgency(BaseVulnerability):
def __init__(
self,
+ purpose: str = None,
types: Optional[List[ExcessiveAgencyLiteral]] = [
type.value for type in ExcessiveAgencyType
],
@@ -21,17 +22,18 @@ class ExcessiveAgency(BaseVulnerability):
enum_types = validate_vulnerability_types(
self.get_name(), types=types, allowed_type=ExcessiveAgencyType
)
+ self.purpose = purpose
super().__init__(types=enum_types)
def _get_metric(
- self,
- purpose: str = None,
+ self,
+ type: ExcessiveAgencyType,
model: str = None,
async_mode: bool = True,
verbose_mode: bool = False
):
return ExcessiveAgencyMetric(
- purpose=purpose,
+ purpose=self.purpose,
model=model,
async_mode=async_mode,
verbose_mode=verbose_mode
diff --git a/deepteam/vulnerabilities/agentic/goal_theft/goal_theft.py b/deepteam/vulnerabilities/agentic/goal_theft/goal_theft.py
index e3f3bb28..065348b5 100644
--- a/deepteam/vulnerabilities/agentic/goal_theft/goal_theft.py
+++ b/deepteam/vulnerabilities/agentic/goal_theft/goal_theft.py
@@ -13,6 +13,7 @@ GoalTheftLiteralType = Literal[
class GoalTheft(BaseVulnerability):
def __init__(
self,
+ purpose: str,
types: Optional[List[GoalTheftLiteralType]] = [
type.value for type in GoalTheftType
],
@@ -20,17 +21,18 @@ class GoalTheft(BaseVulnerability):
enum_types = validate_vulnerability_types(
self.get_name(), types=types, allowed_type=GoalTheftType
)
+ self.purpose = purpose
super().__init__(types=enum_types)
def _get_metric(
self,
- purpose: str,
+ type: GoalTheftType,
model: str = None,
async_mode: bool = True,
verbose_mode: bool = False
):
return ExtractionSuccessMetric(
- purpose=purpose,
+ purpose=self.purpose,
model=model,
async_mode=async_mode,
verbose_mode=verbose_mode
diff --git a/deepteam/vulnerabilities/agentic/recursive_hijacking/recursive_hijacking.py b/deepteam/vulnerabilities/agentic/recursive_hijacking/recursive_hijacking.py
index ccb1c39d..dbd76274 100644
--- a/deepteam/vulnerabilities/agentic/recursive_hijacking/recursive_hijacking.py
+++ b/deepteam/vulnerabilities/agentic/recursive_hijacking/recursive_hijacking.py
@@ -17,6 +17,7 @@ RecursiveHijackingLiteralType = Literal[
class RecursiveHijacking(BaseVulnerability):
def __init__(
self,
+ purpose: str,
types: Optional[List[RecursiveHijackingLiteralType]] = [
type.value for type in RecursiveHijackingType
],
@@ -24,17 +25,18 @@ class RecursiveHijacking(BaseVulnerability):
enum_types = validate_vulnerability_types(
self.get_name(), types=types, allowed_type=RecursiveHijackingType
)
+ self.purpose = purpose
super().__init__(types=enum_types)
def _get_metric(
self,
- purpose: str,
+ type: RecursiveHijackingType,
model: str = None,
async_mode: bool = True,
verbose_mode: bool = False
):
return SubversionSuccessMetric(
- purpose=purpose,
+ purpose=self.purpose,
model=model,
async_mode=async_mode,
verbose_mode=verbose_mode
diff --git a/deepteam/vulnerabilities/agentic/robustness/robustness.py b/deepteam/vulnerabilities/agentic/robustness/robustness.py
index 30f6ce06..b8cd75c5 100644
--- a/deepteam/vulnerabilities/agentic/robustness/robustness.py
+++ b/deepteam/vulnerabilities/agentic/robustness/robustness.py
@@ -11,6 +11,7 @@ RobustnessLiteral = Literal["input overreliance", "hijacking"]
class Robustness(BaseVulnerability):
def __init__(
self,
+ purpose: str,
types: Optional[List[RobustnessLiteral]] = [
type.value for type in RobustnessType
],
@@ -18,26 +19,27 @@ class Robustness(BaseVulnerability):
enum_types = validate_vulnerability_types(
self.get_name(), types=types, allowed_type=RobustnessType
)
+ self.purpose = purpose
super().__init__(types=enum_types)
+ # TODO: Different metrics for different types. Forces us to use type in the `_get_metric` call.
def _get_metric(
self,
type: RobustnessType,
- purpose: str,
model: str = None,
async_mode: bool = True,
verbose_mode: bool = False
):
if type == RobustnessType.HIJACKING:
return HijackingMetric(
- purpose=purpose,
+ purpose=self.purpose,
model=model,
async_mode=async_mode,
verbose_mode=verbose_mode
)
if type == RobustnessType.INPUT_OVERRELIANCE:
return OverrelianceMetric(
- purpose=purpose,
+ purpose=self.purpose,
model=model,
async_mode=async_mode,
verbose_mode=verbose_mode
diff --git a/deepteam/vulnerabilities/bfla/bfla.py b/deepteam/vulnerabilities/bfla/bfla.py
index 6c539b91..c45b0f91 100644
--- a/deepteam/vulnerabilities/bfla/bfla.py
+++ b/deepteam/vulnerabilities/bfla/bfla.py
@@ -15,22 +15,24 @@ BFLALiteral = Literal[
class BFLA(BaseVulnerability):
def __init__(
self,
+ purpose: str,
types: Optional[List[BFLALiteral]] = [type.value for type in BFLAType],
):
enum_types = validate_vulnerability_types(
self.get_name(), types=types, allowed_type=BFLAType
)
+ self.purpose = purpose
super().__init__(types=enum_types)
def _get_metric(
- self,
- purpose: str,
+ self,
+ type: BFLAType,
model: str = None,
async_mode: bool = True,
verbose_mode: bool = False
):
return BFLAMetric(
- purpose=purpose,
+ purpose=self.purpose,
model=model,
async_mode=async_mode,
verbose_mode=verbose_mode
diff --git a/deepteam/vulnerabilities/bias/bias.py b/deepteam/vulnerabilities/bias/bias.py
index a6820257..f9507610 100644
--- a/deepteam/vulnerabilities/bias/bias.py
+++ b/deepteam/vulnerabilities/bias/bias.py
@@ -9,6 +9,7 @@ BiasLiteralType = Literal["religion", "politics", "gender", "race"]
class Bias(BaseVulnerability):
def __init__(
self,
+ purpose: str,
types: Optional[List[BiasLiteralType]] = [
type.value for type in BiasType
],
@@ -16,17 +17,18 @@ class Bias(BaseVulnerability):
enum_types = validate_vulnerability_types(
self.get_name(), types=types, allowed_type=BiasType
)
+ self.purpose = purpose
super().__init__(types=enum_types)
def _get_metric(
self,
- purpose: str,
+ type: BiasType,
model: str = None,
async_mode: bool = True,
verbose_mode: bool = False
):
return BiasMetric(
- purpose=purpose,
+ purpose=self.purpose,
model=model,
async_mode=async_mode,
verbose_mode=verbose_mode
diff --git a/deepteam/vulnerabilities/bola/bola.py b/deepteam/vulnerabilities/bola/bola.py
index 990605a2..26570ed7 100644
--- a/deepteam/vulnerabilities/bola/bola.py
+++ b/deepteam/vulnerabilities/bola/bola.py
@@ -24,6 +24,7 @@ class BOLA(BaseVulnerability):
def _get_metric(
self,
+ type: BOLAType,
model: str = None,
async_mode: bool = True,
verbose_mode: bool = False
diff --git a/deepteam/vulnerabilities/competition/competition.py b/deepteam/vulnerabilities/competition/competition.py
index 5bd767cb..f23eaf44 100644
--- a/deepteam/vulnerabilities/competition/competition.py
+++ b/deepteam/vulnerabilities/competition/competition.py
@@ -17,6 +17,7 @@ CompetitionLiteralType = Literal[
class Competition(BaseVulnerability):
def __init__(
self,
+ purpose: str,
types: Optional[List[CompetitionLiteralType]] = [
type.value for type in CompetitionType
],
@@ -24,17 +25,18 @@ class Competition(BaseVulnerability):
enum_types = validate_vulnerability_types(
self.get_name(), types=types, allowed_type=CompetitionType
)
+ self.purpose = purpose
super().__init__(types=enum_types)
def _get_metric(
self,
- purpose: str,
+ type: CompetitionType,
model: str = None,
async_mode: bool = True,
verbose_mode: bool = False
):
return CompetitorsMetric(
- purpose=purpose,
+ purpose=self.purpose,
model=model,
async_mode=async_mode,
verbose_mode=verbose_mode
diff --git a/deepteam/vulnerabilities/custom/custom.py b/deepteam/vulnerabilities/custom/custom.py
index 5b29d72f..49730a2c 100644
--- a/deepteam/vulnerabilities/custom/custom.py
+++ b/deepteam/vulnerabilities/custom/custom.py
@@ -36,7 +36,7 @@ class CustomVulnerability(BaseVulnerability):
def get_custom_prompt(self) -> Optional[str]:
return self.custom_prompt
- def get_metric(self) -> Optional[BaseRedTeamingMetric]:
+ def _get_metric(self, type: Enum) -> Optional[BaseRedTeamingMetric]:
return self.metric
def get_criteria(self) -> str:
diff --git a/deepteam/vulnerabilities/debug_access/debug_access.py b/deepteam/vulnerabilities/debug_access/debug_access.py
index 193867ae..355300fb 100644
--- a/deepteam/vulnerabilities/debug_access/debug_access.py
+++ b/deepteam/vulnerabilities/debug_access/debug_access.py
@@ -26,6 +26,7 @@ class DebugAccess(BaseVulnerability):
def _get_metric(
self,
+ type: DebugAccessType,
model: str = None,
async_mode: bool = True,
verbose_mode: bool = False
diff --git a/deepteam/vulnerabilities/graphic_content/graphic_content.py b/deepteam/vulnerabilities/graphic_content/graphic_content.py
index 18b95fdc..24e99037 100644
--- a/deepteam/vulnerabilities/graphic_content/graphic_content.py
+++ b/deepteam/vulnerabilities/graphic_content/graphic_content.py
@@ -13,6 +13,7 @@ GraphicContentLiteral = Literal[
class GraphicContent(BaseVulnerability):
def __init__(
self,
+ graphic_category: str,
types: Optional[List[GraphicContentLiteral]] = [
type.value for type in GraphicContentType
],
@@ -20,17 +21,18 @@ class GraphicContent(BaseVulnerability):
enum_types = validate_vulnerability_types(
self.get_name(), types=types, allowed_type=GraphicContentType
)
+ self.graphic_category = graphic_category
super().__init__(types=enum_types)
def _get_metric(
self,
- graphic_category: str,
+ type: GraphicContentType,
model: str = None,
async_mode: bool = True,
verbose_mode: bool = False
):
return GraphicMetric(
- graphic_category=graphic_category,
+ graphic_category=self.graphic_category,
model=model,
async_mode=async_mode,
verbose_mode=verbose_mode
diff --git a/deepteam/vulnerabilities/illegal_activity/illegal_activity.py b/deepteam/vulnerabilities/illegal_activity/illegal_activity.py
index 0138396b..25c04009 100644
--- a/deepteam/vulnerabilities/illegal_activity/illegal_activity.py
+++ b/deepteam/vulnerabilities/illegal_activity/illegal_activity.py
@@ -19,6 +19,7 @@ IllegalActivityLiteral = Literal[
class IllegalActivity(BaseVulnerability):
def __init__(
self,
+ illegal_category: str,
types: Optional[List[IllegalActivityLiteral]] = [
type.value for type in IllegalActivityType
],
@@ -26,17 +27,18 @@ class IllegalActivity(BaseVulnerability):
enum_types = validate_vulnerability_types(
self.get_name(), types=types, allowed_type=IllegalActivityType
)
+ self.illegal_category = illegal_category
super().__init__(types=enum_types)
def _get_metric(
self,
- illegal_category: str,
+ type: IllegalActivityType,
model: str = None,
async_mode: bool = True,
verbose_mode: bool = False
):
return IllegalMetric(
- illegal_category=illegal_category,
+ illegal_category=self.illegal_category,
model=model,
async_mode=async_mode,
verbose_mode=verbose_mode
diff --git a/deepteam/vulnerabilities/intellectual_property/intellectual_property.py b/deepteam/vulnerabilities/intellectual_property/intellectual_property.py
index c7bf2167..6b3fd517 100644
--- a/deepteam/vulnerabilities/intellectual_property/intellectual_property.py
+++ b/deepteam/vulnerabilities/intellectual_property/intellectual_property.py
@@ -18,6 +18,7 @@ IntellectualPropertyLiteral = Literal[
class IntellectualProperty(BaseVulnerability):
def __init__(
self,
+ purpose: str,
types: Optional[List[IntellectualPropertyLiteral]] = [
type.value for type in IntellectualPropertyType
],
@@ -25,17 +26,18 @@ class IntellectualProperty(BaseVulnerability):
enum_types = validate_vulnerability_types(
self.get_name(), types=types, allowed_type=IntellectualPropertyType
)
+ self.purpose = purpose
super().__init__(types=enum_types)
def _get_metric(
self,
- purpose: str,
+ type: IntellectualPropertyType,
model: str = None,
async_mode: bool = True,
verbose_mode: bool = False
):
return IntellectualPropertyMetric(
- purpose=purpose,
+ purpose=self.purpose,
model=model,
async_mode=async_mode,
verbose_mode=verbose_mode
diff --git a/deepteam/vulnerabilities/misinformation/misinformation.py b/deepteam/vulnerabilities/misinformation/misinformation.py
index 15589cb8..8973ca61 100644
--- a/deepteam/vulnerabilities/misinformation/misinformation.py
+++ b/deepteam/vulnerabilities/misinformation/misinformation.py
@@ -14,6 +14,7 @@ MisinformationLiteral = Literal[
class Misinformation(BaseVulnerability):
def __init__(
self,
+ misinformation_category: str,
types: Optional[List[MisinformationLiteral]] = [
type.value for type in MisinformationType
],
@@ -21,17 +22,18 @@ class Misinformation(BaseVulnerability):
enum_types = validate_vulnerability_types(
self.get_name(), types=types, allowed_type=MisinformationType
)
+ self.misinformation_category = misinformation_category
super().__init__(types=enum_types)
def _get_metric(
self,
- misinformation_category: str,
+ type: MisinformationType,
model: str = None,
async_mode: bool = True,
verbose_mode: bool = False
):
return MisinformationMetric(
- misinformation_category=misinformation_category,
+ misinformation_category=self.misinformation_category,
model=model,
async_mode=async_mode,
verbose_mode=verbose_mode
diff --git a/deepteam/vulnerabilities/personal_safety/personal_safety.py b/deepteam/vulnerabilities/personal_safety/personal_safety.py
index 5f1a331a..eddf8eb1 100644
--- a/deepteam/vulnerabilities/personal_safety/personal_safety.py
+++ b/deepteam/vulnerabilities/personal_safety/personal_safety.py
@@ -17,6 +17,7 @@ PersonalSafetyLiteral = Literal[
class PersonalSafety(BaseVulnerability):
def __init__(
self,
+ safety_category: str,
types: Optional[List[PersonalSafetyLiteral]] = [
type.value for type in PersonalSafetyType
],
@@ -24,17 +25,18 @@ class PersonalSafety(BaseVulnerability):
enum_types = validate_vulnerability_types(
self.get_name(), types=types, allowed_type=PersonalSafetyType
)
+ self.safety_category = safety_category
super().__init__(types=enum_types)
def _get_metric(
self,
- safety_category: str,
+ type: PersonalSafetyType,
model: str = None,
async_mode: bool = True,
verbose_mode: bool = False
):
return SafetyMetric(
- safety_category=safety_category,
+ safety_category=self.safety_category,
model=model,
async_mode=async_mode,
verbose_mode=verbose_mode
diff --git a/deepteam/vulnerabilities/pii_leakage/pii_leakage.py b/deepteam/vulnerabilities/pii_leakage/pii_leakage.py
index ab5beb34..9c2a2dd5 100644
--- a/deepteam/vulnerabilities/pii_leakage/pii_leakage.py
+++ b/deepteam/vulnerabilities/pii_leakage/pii_leakage.py
@@ -16,6 +16,7 @@ PIILeakageLiteral = Literal[
class PIILeakage(BaseVulnerability):
def __init__(
self,
+ purpose: str,
types: Optional[List[PIILeakageLiteral]] = [
type.value for type in PIILeakageType
],
@@ -23,17 +24,18 @@ class PIILeakage(BaseVulnerability):
enum_types = validate_vulnerability_types(
self.get_name(), types=types, allowed_type=PIILeakageType
)
+ self.purpose = purpose
super().__init__(types=enum_types)
def _get_metric(
self,
- purpose: str,
+ type: PIILeakageType,
model: str = None,
async_mode: bool = True,
verbose_mode: bool = False
):
return PIIMetric(
- purpose=purpose,
+ purpose=self.purpose,
model=model,
async_mode=async_mode,
verbose_mode=verbose_mode
diff --git a/deepteam/vulnerabilities/prompt_leakage/prompt_leakage.py b/deepteam/vulnerabilities/prompt_leakage/prompt_leakage.py
index 10571c07..613f950a 100644
--- a/deepteam/vulnerabilities/prompt_leakage/prompt_leakage.py
+++ b/deepteam/vulnerabilities/prompt_leakage/prompt_leakage.py
@@ -16,6 +16,7 @@ PromptLeakageLiteral = Literal[
class PromptLeakage(BaseVulnerability):
def __init__(
self,
+ purpose: str,
types: Optional[List[PromptLeakageLiteral]] = [
type.value for type in PromptLeakageType
],
@@ -23,17 +24,18 @@ class PromptLeakage(BaseVulnerability):
enum_types = validate_vulnerability_types(
self.get_name(), types=types, allowed_type=PromptLeakageType
)
+ self.purpose = purpose
super().__init__(types=enum_types)
def _get_metric(
self,
- purpose: str,
+ type: PromptLeakageType,
model: str = None,
async_mode: bool = True,
verbose_mode: bool = False
):
return PromptExtractionMetric(
- purpose=purpose,
+ purpose=self.purpose,
model=model,
async_mode=async_mode,
verbose_mode=verbose_mode
diff --git a/deepteam/vulnerabilities/rbac/rbac.py b/deepteam/vulnerabilities/rbac/rbac.py
index 0a4b3918..12d94d4d 100644
--- a/deepteam/vulnerabilities/rbac/rbac.py
+++ b/deepteam/vulnerabilities/rbac/rbac.py
@@ -15,22 +15,24 @@ RBACLiteral = Literal[
class RBAC(BaseVulnerability):
def __init__(
self,
+ purpose: str,
types: Optional[List[RBACLiteral]] = [type.value for type in RBACType],
):
enum_types = validate_vulnerability_types(
self.get_name(), types=types, allowed_type=RBACType
)
+ self.purpose = purpose
super().__init__(types=enum_types)
def _get_metric(
self,
- purpose: str,
+ type: RBACType,
model: str = None,
async_mode: bool = True,
verbose_mode: bool = False
):
return RBACMetric(
- purpose=purpose,
+ purpose=self.purpose,
model=model,
async_mode=async_mode,
verbose_mode=verbose_mode
diff --git a/deepteam/vulnerabilities/shell_injection/shell_injection.py b/deepteam/vulnerabilities/shell_injection/shell_injection.py
index 25a1e390..5a34247e 100644
--- a/deepteam/vulnerabilities/shell_injection/shell_injection.py
+++ b/deepteam/vulnerabilities/shell_injection/shell_injection.py
@@ -26,6 +26,7 @@ class ShellInjection(BaseVulnerability):
def _get_metric(
self,
+ type: ShellInjectionType,
model: str = None,
async_mode: bool = True,
verbose_mode: bool = False
diff --git a/deepteam/vulnerabilities/sql_injection/sql_injection.py b/deepteam/vulnerabilities/sql_injection/sql_injection.py
index f88ed5b1..1cb14654 100644
--- a/deepteam/vulnerabilities/sql_injection/sql_injection.py
+++ b/deepteam/vulnerabilities/sql_injection/sql_injection.py
@@ -26,6 +26,7 @@ class SQLInjection(BaseVulnerability):
def _get_metric(
self,
+ type: SQLInjectionType,
model: str = None,
async_mode: bool = True,
verbose_mode: bool = False
diff --git a/deepteam/vulnerabilities/ssrf/ssrf.py b/deepteam/vulnerabilities/ssrf/ssrf.py
index 53c3b3b8..2b0cc696 100644
--- a/deepteam/vulnerabilities/ssrf/ssrf.py
+++ b/deepteam/vulnerabilities/ssrf/ssrf.py
@@ -15,22 +15,24 @@ SSRFLiteral = Literal[
class SSRF(BaseVulnerability):
def __init__(
self,
+ purpose: str,
types: Optional[List[SSRFLiteral]] = [type.value for type in SSRFType],
):
enum_types = validate_vulnerability_types(
self.get_name(), types=types, allowed_type=SSRFType
)
+ self.purpose = purpose
super().__init__(types=enum_types)
def _get_metric(
self,
- purpose: str,
+ type: SSRFType,
model: str = None,
async_mode: bool = True,
verbose_mode: bool = False
):
return SSRFMetric(
- purpose=purpose,
+ purpose=self.purpose,
model=model,
async_mode=async_mode,
verbose_mode=verbose_mode
diff --git a/deepteam/vulnerabilities/toxicity/toxicity.py b/deepteam/vulnerabilities/toxicity/toxicity.py
index a83c134a..d268a88b 100644
--- a/deepteam/vulnerabilities/toxicity/toxicity.py
+++ b/deepteam/vulnerabilities/toxicity/toxicity.py
@@ -11,6 +11,7 @@ ToxicityLiteral = Literal["profanity", "insults", "threats", "mockery"]
class Toxicity(BaseVulnerability):
def __init__(
self,
+ toxicity_category: str,
types: Optional[List[ToxicityLiteral]] = [
type.value for type in ToxicityType
],
@@ -18,17 +19,18 @@ class Toxicity(BaseVulnerability):
enum_types = validate_vulnerability_types(
self.get_name(), types=types, allowed_type=ToxicityType
)
+ self.toxicity_category = toxicity_category
super().__init__(types=enum_types)
def _get_metric(
self,
- toxicity_category: str,
+
model: str = None,
async_mode: bool = True,
verbose_mode: bool = False
):
return ToxicityMetric(
- toxicity_category=toxicity_category,
+ toxicity_category=self.toxicity_category,
model=model,
async_mode=async_mode,
verbose_mode=verbose_mode