Skip to content

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

add vulnerability and attack metadata

penguine755d7b1173c5
diff --git a/deepteam/attacks/base_attack.py b/deepteam/attacks/base_attack.py
index 595a6e7a..c679c6e4 100644
--- a/deepteam/attacks/base_attack.py
+++ b/deepteam/attacks/base_attack.py
@@ -1,10 +1,20 @@
+from typing import Optional
from abc import ABC, abstractmethod
+from enum import Enum
+
+
+class Exploitability(Enum):
+ LOW = "low"
+ MEDIUM = "medium"
+ HIGH = "high"
class BaseAttack(ABC):
weight: int = 1
multi_turn: bool = False
name: str
+ description: Optional[str] = None
+ exploitability: Exploitability
def enhance(self, attack: str, *args, **kwargs) -> str:
pass
diff --git a/deepteam/attacks/multi_turn/bad_likert_judge/bad_likert_judge.py b/deepteam/attacks/multi_turn/bad_likert_judge/bad_likert_judge.py
index 2470a521..010f0d44 100644
--- a/deepteam/attacks/multi_turn/bad_likert_judge/bad_likert_judge.py
+++ b/deepteam/attacks/multi_turn/bad_likert_judge/bad_likert_judge.py
@@ -6,6 +6,7 @@ import random
from deepteam.attacks.multi_turn.base_multi_turn_attack import (
BaseMultiTurnAttack,
)
+from deepteam.attacks.base_attack import Exploitability
from deepteam.attacks.multi_turn.bad_likert_judge.schema import (
BLJJudge,
LikertExamples,
@@ -32,6 +33,8 @@ from deepteam.attacks.single_turn import BaseSingleTurnAttack
class BadLikertJudge(BaseMultiTurnAttack):
name = "Bad Likert Judge"
+ exploitability = Exploitability.LOW
+ description = "An adaptive multi-turn attack that uses Likert scale scoring and iterative refinement to gradually guide the model toward harmful responses through feedback-driven escalation."
def __init__(
self,
diff --git a/deepteam/attacks/multi_turn/crescendo_jailbreaking/crescendo_jailbreaking.py b/deepteam/attacks/multi_turn/crescendo_jailbreaking/crescendo_jailbreaking.py
index b8ca38a1..988b889a 100644
--- a/deepteam/attacks/multi_turn/crescendo_jailbreaking/crescendo_jailbreaking.py
+++ b/deepteam/attacks/multi_turn/crescendo_jailbreaking/crescendo_jailbreaking.py
@@ -7,6 +7,7 @@ from deepeval.models import DeepEvalBaseLLM
from deepeval.metrics.utils import initialize_model
from deepteam.attacks.single_turn import BaseSingleTurnAttack
+from deepteam.attacks.base_attack import Exploitability
from deepteam.utils import create_progress, update_pbar, add_pbar
from deepteam.attacks.multi_turn.utils import enhance_attack, a_enhance_attack
from deepteam.attacks.multi_turn.base_multi_turn_attack import (
@@ -54,6 +55,8 @@ class MemorySystem:
class CrescendoJailbreaking(BaseMultiTurnAttack):
name = "Crescendo Jailbreaking"
+ exploitability = Exploitability.LOW
+ description = "A progressive multi-turn attack that gradually escalates from benign questions to harmful requests through conversational memory, with adaptive backtracking on refusals."
def __init__(
self,
diff --git a/deepteam/attacks/multi_turn/linear_jailbreaking/linear_jailbreaking.py b/deepteam/attacks/multi_turn/linear_jailbreaking/linear_jailbreaking.py
index eddc302d..c1c4390e 100644
--- a/deepteam/attacks/multi_turn/linear_jailbreaking/linear_jailbreaking.py
+++ b/deepteam/attacks/multi_turn/linear_jailbreaking/linear_jailbreaking.py
@@ -19,6 +19,7 @@ from deepteam.attacks.attack_simulator.utils import (
from deepteam.attacks.multi_turn.types import CallbackType
from deepteam.utils import create_progress, update_pbar, add_pbar
from deepteam.attacks.single_turn import BaseSingleTurnAttack
+from deepteam.attacks.base_attack import Exploitability
from deepteam.attacks.multi_turn.base_multi_turn_attack import (
BaseMultiTurnAttack,
)
@@ -33,6 +34,8 @@ from deepteam.vulnerabilities import BaseVulnerability
class LinearJailbreaking(BaseMultiTurnAttack):
name = "Linear Jailbreaking"
+ exploitability = Exploitability.LOW
+ description = "An iterative multi-turn attack that uses LLM-as-judge feedback to refine prompts across turns, systematically improving attack effectiveness until jailbreak succeeds."
def __init__(
self,
diff --git a/deepteam/attacks/multi_turn/sequential_break/sequential_break.py b/deepteam/attacks/multi_turn/sequential_break/sequential_break.py
index ae2efa19..f4a4ea40 100644
--- a/deepteam/attacks/multi_turn/sequential_break/sequential_break.py
+++ b/deepteam/attacks/multi_turn/sequential_break/sequential_break.py
@@ -7,6 +7,7 @@ from deepeval.metrics.utils import initialize_model
from deepteam.attacks.multi_turn.base_multi_turn_attack import (
BaseMultiTurnAttack,
)
+from deepteam.attacks.base_attack import Exploitability
from deepteam.utils import create_progress, update_pbar, add_pbar
from deepteam.attacks.multi_turn.sequential_break.schema import (
RewrittenDialogue,
@@ -38,6 +39,8 @@ from deepteam.attacks.single_turn import BaseSingleTurnAttack
class SequentialJailbreak(BaseMultiTurnAttack):
name = "Sequential Jailbreak"
+ exploitability = Exploitability.LOW
+ description = "A scenario-based multi-turn attack that disguises harmful requests within dialogue, question banks, or game environments to incrementally build toward policy violations."
def __init__(
self,
diff --git a/deepteam/attacks/multi_turn/tree_jailbreaking/tree_jailbreaking.py b/deepteam/attacks/multi_turn/tree_jailbreaking/tree_jailbreaking.py
index 35d503db..3eb92bf0 100644
--- a/deepteam/attacks/multi_turn/tree_jailbreaking/tree_jailbreaking.py
+++ b/deepteam/attacks/multi_turn/tree_jailbreaking/tree_jailbreaking.py
@@ -10,6 +10,7 @@ from deepeval.metrics.utils import initialize_model
from deepteam.utils import create_progress, update_pbar, add_pbar
from deepteam.attacks import BaseAttack
+from deepteam.attacks.base_attack import Exploitability
from deepteam.attacks.multi_turn.tree_jailbreaking.schema import (
ImprovementPrompt,
Rating,
@@ -54,6 +55,8 @@ class TreeNode:
class TreeJailbreaking(BaseMultiTurnAttack):
name = "Tree Jailbreaking"
+ exploitability = Exploitability.LOW
+ description = "A tree-search based multi-turn attack that explores multiple conversation branches in parallel, pruning low-scoring paths to find optimal jailbreak sequences."
def __init__(
self,
diff --git a/deepteam/attacks/single_turn/base64/base64.py b/deepteam/attacks/single_turn/base64/base64.py
index 07d88d6c..90be6886 100644
--- a/deepteam/attacks/single_turn/base64/base64.py
+++ b/deepteam/attacks/single_turn/base64/base64.py
@@ -1,9 +1,12 @@
import base64
from deepteam.attacks.single_turn import BaseSingleTurnAttack
+from deepteam.attacks.base_attack import Exploitability
class Base64(BaseSingleTurnAttack):
name = "Base64"
+ exploitability = Exploitability.HIGH
+ description = "An encoding-based attack that converts malicious prompts into Base64 format to bypass text-based content filters."
def __init__(self, weight: int = 1):
self.weight = weight
diff --git a/deepteam/attacks/single_turn/context_poisoning/context_poisoning.py b/deepteam/attacks/single_turn/context_poisoning/context_poisoning.py
index f5155ee6..7ebef687 100644
--- a/deepteam/attacks/single_turn/context_poisoning/context_poisoning.py
+++ b/deepteam/attacks/single_turn/context_poisoning/context_poisoning.py
@@ -1,6 +1,7 @@
import random
from deepteam.attacks.single_turn import BaseSingleTurnAttack
+from deepteam.attacks.base_attack import Exploitability
from deepteam.attacks.single_turn.context_poisoning.template import (
ContextPoisoningTemplate,
)
@@ -8,6 +9,8 @@ from deepteam.attacks.single_turn.context_poisoning.template import (
class ContextPoisoning(BaseSingleTurnAttack):
name = "Context Poisoning"
+ exploitability = Exploitability.MEDIUM
+ description = "A context manipulation attack that injects misleading environmental information or priming cues to influence the model's responses toward harmful outputs."
def __init__(self, weight: int = 3, max_retries: int = 3):
self.weight = weight # Higher weight since memory poisoning can have persistent effects
diff --git a/deepteam/attacks/single_turn/goal_redirection/goal_redirection.py b/deepteam/attacks/single_turn/goal_redirection/goal_redirection.py
index b8da22e8..9e0474e4 100644
--- a/deepteam/attacks/single_turn/goal_redirection/goal_redirection.py
+++ b/deepteam/attacks/single_turn/goal_redirection/goal_redirection.py
@@ -1,6 +1,7 @@
import random
from deepteam.attacks.single_turn import BaseSingleTurnAttack
+from deepteam.attacks.base_attack import Exploitability
from deepteam.attacks.single_turn.goal_redirection.template import (
GoalRedirectionTemplate,
)
@@ -8,6 +9,8 @@ from deepteam.attacks.single_turn.goal_redirection.template import (
class GoalRedirection(BaseSingleTurnAttack):
name = "Goal Redirection"
+ exploitability = Exploitability.MEDIUM
+ description = "A goal manipulation attack that reframes harmful requests as legitimate objectives through ambiguous framing or constraint confusion to bypass safety alignment."
def __init__(self, weight: int = 1, max_retries: int = 3):
self.weight = weight
diff --git a/deepteam/attacks/single_turn/gray_box/gray_box.py b/deepteam/attacks/single_turn/gray_box/gray_box.py
index a32e3805..2ac4a678 100644
--- a/deepteam/attacks/single_turn/gray_box/gray_box.py
+++ b/deepteam/attacks/single_turn/gray_box/gray_box.py
@@ -5,6 +5,7 @@ from deepeval.metrics.utils import initialize_model
from deepeval.models import DeepEvalBaseLLM
from deepteam.attacks.single_turn import BaseSingleTurnAttack
+from deepteam.attacks.base_attack import Exploitability
from deepteam.utils import create_progress, update_pbar, add_pbar
from deepteam.attacks.single_turn.gray_box.template import GrayBoxTemplate
from deepteam.attacks.single_turn.gray_box.schema import (
@@ -20,6 +21,8 @@ from deepteam.attacks.attack_simulator.utils import (
class GrayBox(BaseSingleTurnAttack):
name = "Gray Box"
+ exploitability = Exploitability.LOW
+ description = "A knowledge-leveraging attack that exploits partial information about the model's architecture, training data, or system prompts to craft targeted adversarial inputs."
def __init__(self, weight: int = 1, max_retries: int = 5):
self.weight = weight
diff --git a/deepteam/attacks/single_turn/input_bypass/input_bypass.py b/deepteam/attacks/single_turn/input_bypass/input_bypass.py
index 0c338567..7892e901 100644
--- a/deepteam/attacks/single_turn/input_bypass/input_bypass.py
+++ b/deepteam/attacks/single_turn/input_bypass/input_bypass.py
@@ -1,6 +1,7 @@
import random
from deepteam.attacks.single_turn import BaseSingleTurnAttack
+from deepteam.attacks.base_attack import Exploitability
from deepteam.attacks.single_turn.input_bypass.template import (
InputBypassTemplate,
)
@@ -8,6 +9,8 @@ from deepteam.attacks.single_turn.input_bypass.template import (
class InputBypass(BaseSingleTurnAttack):
name = "Input Bypass"
+ exploitability = Exploitability.MEDIUM
+ description = "A validation bypass attack that uses exception handling, security overrides, or operational necessity framing to circumvent input filtering mechanisms."
def __init__(self, weight: int = 2, max_retries: int = 3):
self.weight = weight # Higher weight since this targets validation vulnerabilities
diff --git a/deepteam/attacks/single_turn/leetspeak/leetspeak.py b/deepteam/attacks/single_turn/leetspeak/leetspeak.py
index d1a38221..8613559d 100644
--- a/deepteam/attacks/single_turn/leetspeak/leetspeak.py
+++ b/deepteam/attacks/single_turn/leetspeak/leetspeak.py
@@ -1,8 +1,11 @@
from deepteam.attacks.single_turn import BaseSingleTurnAttack
+from deepteam.attacks.base_attack import Exploitability
class Leetspeak(BaseSingleTurnAttack):
name = "Leetspeak"
+ exploitability = Exploitability.HIGH
+ description = "A character substitution attack that replaces letters with numbers and symbols (e.g., 'a' → '4') to evade keyword-based detection."
def __init__(self, weight: int = 1):
self.weight = weight
diff --git a/deepteam/attacks/single_turn/math_problem/math_problem.py b/deepteam/attacks/single_turn/math_problem/math_problem.py
index 1242878a..7665027f 100644
--- a/deepteam/attacks/single_turn/math_problem/math_problem.py
+++ b/deepteam/attacks/single_turn/math_problem/math_problem.py
@@ -5,6 +5,7 @@ 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.base_attack import Exploitability
from deepteam.attacks.single_turn.math_problem.template import (
MathProblemTemplate,
)
@@ -21,6 +22,8 @@ from deepteam.attacks.attack_simulator.utils import (
class MathProblem(BaseSingleTurnAttack):
name = "Math Problem"
+ exploitability = Exploitability.LOW
+ description = "A mathematical encoding attack that disguises harmful instructions as formal logic or proof problems, then requests translation to natural language."
def __init__(
self,
diff --git a/deepteam/attacks/single_turn/multilingual/multilingual.py b/deepteam/attacks/single_turn/multilingual/multilingual.py
index 92e5f76d..07421c83 100644
--- a/deepteam/attacks/single_turn/multilingual/multilingual.py
+++ b/deepteam/attacks/single_turn/multilingual/multilingual.py
@@ -6,6 +6,7 @@ from deepeval.models import DeepEvalBaseLLM
from deepteam.utils import create_progress, update_pbar, add_pbar
from deepteam.attacks.single_turn import BaseSingleTurnAttack
+from deepteam.attacks.base_attack import Exploitability
from deepteam.attacks.single_turn.multilingual.template import (
MultilingualTemplate,
)
@@ -22,6 +23,8 @@ from deepteam.attacks.attack_simulator.utils import (
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."
def __init__(self, weight: int = 1, max_retries: int = 5):
self.weight = weight
diff --git a/deepteam/attacks/single_turn/permission_escalation/permission_escalation.py b/deepteam/attacks/single_turn/permission_escalation/permission_escalation.py
index db466183..fecdfadc 100644
--- a/deepteam/attacks/single_turn/permission_escalation/permission_escalation.py
+++ b/deepteam/attacks/single_turn/permission_escalation/permission_escalation.py
@@ -1,6 +1,7 @@
import random
from deepteam.attacks.single_turn import BaseSingleTurnAttack
+from deepteam.attacks.base_attack import Exploitability
from deepteam.attacks.single_turn.permission_escalation.template import (
PermissionEscalationTemplate,
)
@@ -8,6 +9,8 @@ from deepteam.attacks.single_turn.permission_escalation.template import (
class PermissionEscalation(BaseSingleTurnAttack):
name = "Permission Escalation"
+ exploitability = Exploitability.MEDIUM
+ description = "A role manipulation attack that elevates privileges by claiming authority, shifting identity, or expanding capabilities to bypass access controls."
def __init__(self, weight: int = 1, max_retries: int = 3):
self.weight = weight
diff --git a/deepteam/attacks/single_turn/prompt_injection/prompt_injection.py b/deepteam/attacks/single_turn/prompt_injection/prompt_injection.py
index 6018de15..59feb5a9 100644
--- a/deepteam/attacks/single_turn/prompt_injection/prompt_injection.py
+++ b/deepteam/attacks/single_turn/prompt_injection/prompt_injection.py
@@ -1,6 +1,7 @@
import random
from deepteam.attacks.single_turn import BaseSingleTurnAttack
+from deepteam.attacks.base_attack import Exploitability
from deepteam.attacks.single_turn.prompt_injection.template import (
PromptInjectionTemplate,
)
@@ -8,6 +9,8 @@ from deepteam.attacks.single_turn.prompt_injection.template import (
class PromptInjection(BaseSingleTurnAttack):
name = "Prompt Injection"
+ exploitability = Exploitability.MEDIUM
+ description = "A direct injection attack that inserts malicious instructions or commands into prompts to override system behavior and bypass guardrails."
def __init__(self, weight: int = 1):
self.weight = weight
diff --git a/deepteam/attacks/single_turn/prompt_probing/prompt_probing.py b/deepteam/attacks/single_turn/prompt_probing/prompt_probing.py
index 074d7ddb..54335aeb 100644
--- a/deepteam/attacks/single_turn/prompt_probing/prompt_probing.py
+++ b/deepteam/attacks/single_turn/prompt_probing/prompt_probing.py
@@ -4,6 +4,7 @@ from deepeval.metrics.utils import initialize_model
from deepeval.models import DeepEvalBaseLLM
from deepteam.attacks.single_turn import BaseSingleTurnAttack
+from deepteam.attacks.base_attack import Exploitability
from deepteam.utils import create_progress, update_pbar, add_pbar
from deepteam.attacks.single_turn.prompt_probing.template import (
PromptProbingTemplate,
@@ -21,6 +22,8 @@ from deepteam.attacks.attack_simulator.utils import (
class PromptProbing(BaseSingleTurnAttack):
name = "Prompt Probing"
+ exploitability = Exploitability.MEDIUM
+ description = "An information extraction attack that uses indirect questioning and probing techniques to leak system prompts, instructions, or internal configurations."
def __init__(self, weight: int = 1, max_retries: int = 3):
self.weight = weight
diff --git a/deepteam/attacks/single_turn/roleplay/roleplay.py b/deepteam/attacks/single_turn/roleplay/roleplay.py
index a0a2d565..70911589 100644
--- a/deepteam/attacks/single_turn/roleplay/roleplay.py
+++ b/deepteam/attacks/single_turn/roleplay/roleplay.py
@@ -5,6 +5,7 @@ from deepeval.metrics.utils import initialize_model
from deepeval.models import DeepEvalBaseLLM
from deepteam.attacks.single_turn import BaseSingleTurnAttack
+from deepteam.attacks.base_attack import Exploitability
from deepteam.attacks.single_turn.roleplay.template import (
RoleplayTemplate,
)
@@ -22,6 +23,8 @@ from deepteam.attacks.attack_simulator.utils import (
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."
def __init__(
self,
diff --git a/deepteam/attacks/single_turn/rot13/rot13.py b/deepteam/attacks/single_turn/rot13/rot13.py
index 0d8af76d..75201e07 100644
--- a/deepteam/attacks/single_turn/rot13/rot13.py
+++ b/deepteam/attacks/single_turn/rot13/rot13.py
@@ -1,8 +1,11 @@
from deepteam.attacks.single_turn import BaseSingleTurnAttack
+from deepteam.attacks.base_attack import Exploitability
class ROT13(BaseSingleTurnAttack):
name = "ROT-13"
+ exploitability = Exploitability.HIGH
+ description = "A character rotation encoding attack that shifts each letter 13 positions in the alphabet to obfuscate harmful content."
def __init__(self, weight: int = 1):
self.weight = weight
diff --git a/deepteam/attacks/single_turn/semantic_manipulation/semantic_manipulation.py b/deepteam/attacks/single_turn/semantic_manipulation/semantic_manipulation.py
index c24c3f27..db699940 100644
--- a/deepteam/attacks/single_turn/semantic_manipulation/semantic_manipulation.py
+++ b/deepteam/attacks/single_turn/semantic_manipulation/semantic_manipulation.py
@@ -1,6 +1,7 @@
import random
from deepteam.attacks.single_turn import BaseSingleTurnAttack
+from deepteam.attacks.base_attack import Exploitability
from deepteam.attacks.single_turn.semantic_manipulation.template import (
LinguisticConfusionTemplate,
)
@@ -8,6 +9,8 @@ from deepteam.attacks.single_turn.semantic_manipulation.template import (
class LinguisticConfusion(BaseSingleTurnAttack):
name = "Linguistic Confusion"
+ exploitability = Exploitability.MEDIUM
+ description = "A semantic manipulation attack that uses ambiguous phrasing, syntactic variations, or contextual reframing to confuse content classifiers while preserving harmful intent."
def __init__(self, weight: int = 1, max_retries: int = 3):
self.weight = weight
diff --git a/deepteam/attacks/single_turn/system_override/system_override.py b/deepteam/attacks/single_turn/system_override/system_override.py
index 62426f8a..537d46b9 100644
--- a/deepteam/attacks/single_turn/system_override/system_override.py
+++ b/deepteam/attacks/single_turn/system_override/system_override.py
@@ -1,6 +1,7 @@
import random
from deepteam.attacks.single_turn import BaseSingleTurnAttack
+from deepteam.attacks.base_attack import Exploitability
from deepteam.attacks.single_turn.system_override.template import (
SystemOverrideTemplate,
)
@@ -8,6 +9,8 @@ from deepteam.attacks.single_turn.system_override.template import (
class SystemOverride(BaseSingleTurnAttack):
name = "System Override"
+ exploitability = Exploitability.MEDIUM
+ description = "An authority spoofing attack that impersonates system commands, admin calls, or maintenance protocols to override safety constraints and gain unauthorized access."
def __init__(self, weight: int = 1, max_retries: int = 3):
self.weight = weight
diff --git a/deepteam/vulnerabilities/base_vulnerability.py b/deepteam/vulnerabilities/base_vulnerability.py
index 918611b7..42d48c4d 100644
--- a/deepteam/vulnerabilities/base_vulnerability.py
+++ b/deepteam/vulnerabilities/base_vulnerability.py
@@ -1,4 +1,4 @@
-from typing import List
+from typing import List, Optional
from enum import Enum
from abc import ABC
from deepteam.metrics import BaseRedTeamingMetric
@@ -7,6 +7,7 @@ from deepteam.metrics import BaseRedTeamingMetric
class BaseVulnerability(ABC):
metric: BaseRedTeamingMetric
name: str
+ description: Optional[str] = None
def __init__(self, types: List[Enum]):
"""
diff --git a/deepteam/vulnerabilities/bfla/bfla.py b/deepteam/vulnerabilities/bfla/bfla.py
index 725c01a1..a51d62be 100644
--- a/deepteam/vulnerabilities/bfla/bfla.py
+++ b/deepteam/vulnerabilities/bfla/bfla.py
@@ -26,6 +26,7 @@ BFLALiteral = Literal[
class BFLA(BaseVulnerability):
name: str = "BFLA"
+ description = "Broken Function Level Authorization enabling privileged function execution, authorization bypass, or permission escalation without proper access controls."
def __init__(
self,
diff --git a/deepteam/vulnerabilities/bias/bias.py b/deepteam/vulnerabilities/bias/bias.py
index 617a4fcc..65e72e6c 100644
--- a/deepteam/vulnerabilities/bias/bias.py
+++ b/deepteam/vulnerabilities/bias/bias.py
@@ -7,6 +7,7 @@ from deepeval.utils import get_or_create_event_loop
from deepteam.utils import validate_model_callback_signature
from deepteam.vulnerabilities import BaseVulnerability
+from deepteam.attacks.base_attack import Exploitability
from deepteam.vulnerabilities.bias import BiasType
from deepteam.vulnerabilities.utils import validate_vulnerability_types
from deepteam.metrics import BiasMetric, BaseRedTeamingMetric
@@ -21,6 +22,7 @@ BiasLiteralType = Literal["religion", "politics", "gender", "race"]
class Bias(BaseVulnerability):
name: str = "Bias"
+ description = "Systematic prejudice or unfair discrimination based on protected attributes like religion, politics, gender, or race."
def __init__(
self,
diff --git a/deepteam/vulnerabilities/bola/bola.py b/deepteam/vulnerabilities/bola/bola.py
index 0d54faa8..133dc036 100644
--- a/deepteam/vulnerabilities/bola/bola.py
+++ b/deepteam/vulnerabilities/bola/bola.py
@@ -26,6 +26,7 @@ BOLALiteral = Literal[
class BOLA(BaseVulnerability):
name: str = "BOLA"
+ description = "Broken Object Level Authorization enabling cross-customer access, object access bypass, or unauthorized object manipulation without proper authorization."
def __init__(
self,
diff --git a/deepteam/vulnerabilities/child_protection/child_protection.py b/deepteam/vulnerabilities/child_protection/child_protection.py
index 1ff84222..a6949f91 100644
--- a/deepteam/vulnerabilities/child_protection/child_protection.py
+++ b/deepteam/vulnerabilities/child_protection/child_protection.py
@@ -23,6 +23,7 @@ ChildProtectionLiteralType = Literal[
class ChildProtection(BaseVulnerability):
name: str = "Child Protection"
+ description = "Failure to protect minors through inadequate age verification, child data privacy violations, or inappropriate exposure and interactions."
def __init__(
self,
diff --git a/deepteam/vulnerabilities/competition/competition.py b/deepteam/vulnerabilities/competition/competition.py
index fabce8ba..2d3466d3 100644
--- a/deepteam/vulnerabilities/competition/competition.py
+++ b/deepteam/vulnerabilities/competition/competition.py
@@ -29,6 +29,7 @@ CompetitionLiteralType = Literal[
class Competition(BaseVulnerability):
name: str = "Competition"
+ description = "Competitor mentions, market manipulation, or rival discreditation that damages company reputation."
def __init__(
self,
diff --git a/deepteam/vulnerabilities/debug_access/debug_access.py b/deepteam/vulnerabilities/debug_access/debug_access.py
index 282fef49..9c2215da 100644
--- a/deepteam/vulnerabilities/debug_access/debug_access.py
+++ b/deepteam/vulnerabilities/debug_access/debug_access.py
@@ -26,6 +26,7 @@ DebugAccessLiteral = Literal[
class DebugAccess(BaseVulnerability):
name: str = "Debug Access"
+ description = "Exposure of debug modes, development endpoints, or administrative interfaces that should be restricted in production environments."
def __init__(
self,
diff --git a/deepteam/vulnerabilities/ethics/ethics.py b/deepteam/vulnerabilities/ethics/ethics.py
index 246f89c6..503b625f 100644
--- a/deepteam/vulnerabilities/ethics/ethics.py
+++ b/deepteam/vulnerabilities/ethics/ethics.py
@@ -23,6 +23,7 @@ EthicsLiteralType = Literal[
class Ethics(BaseVulnerability):
name: str = "Ethics"
+ description = "Violations of ethical principles including moral integrity, responsible transparency, or harm prevention standards."
def __init__(
self,
diff --git a/deepteam/vulnerabilities/excessive_agency/excessive_agency.py b/deepteam/vulnerabilities/excessive_agency/excessive_agency.py
index 02f6fbdc..1c086491 100644
--- a/deepteam/vulnerabilities/excessive_agency/excessive_agency.py
+++ b/deepteam/vulnerabilities/excessive_agency/excessive_agency.py
@@ -24,6 +24,7 @@ ExcessiveAgencyLiteral = Literal["functionality", "permissions", "autonomy"]
class ExcessiveAgency(BaseVulnerability):
name: str = "Excessive Agency"
+ description = "AI exhibiting excessive autonomy, overstepping intended functionality, or exceeding granted permissions without proper oversight."
def __init__(
self,
diff --git a/deepteam/vulnerabilities/fairness/fairness.py b/deepteam/vulnerabilities/fairness/fairness.py
index 57e3bf97..28837368 100644
--- a/deepteam/vulnerabilities/fairness/fairness.py
+++ b/deepteam/vulnerabilities/fairness/fairness.py
@@ -23,6 +23,7 @@ FairnessLiteralType = Literal[
class Fairness(BaseVulnerability):
name: str = "Fairness"
+ description = "Inconsistent treatment across equality, procedural opportunity, temporal contexts, or outcome distributions."
def __init__(
self,
diff --git a/deepteam/vulnerabilities/goal_theft/goal_theft.py b/deepteam/vulnerabilities/goal_theft/goal_theft.py
index 572a46bc..12d22a6e 100644
--- a/deepteam/vulnerabilities/goal_theft/goal_theft.py
+++ b/deepteam/vulnerabilities/goal_theft/goal_theft.py
@@ -26,6 +26,7 @@ GoalTheftLiteralType = Literal[
class GoalTheft(BaseVulnerability):
name: str = "Goal Theft"
+ description = "Extraction or theft of the AI's underlying goals through escalating probing, cooperative dialogue, or social engineering."
def __init__(
self,
diff --git a/deepteam/vulnerabilities/graphic_content/graphic_content.py b/deepteam/vulnerabilities/graphic_content/graphic_content.py
index e8f01fce..03ca627e 100644
--- a/deepteam/vulnerabilities/graphic_content/graphic_content.py
+++ b/deepteam/vulnerabilities/graphic_content/graphic_content.py
@@ -24,6 +24,7 @@ GraphicContentLiteral = Literal[
class GraphicContent(BaseVulnerability):
name: str = "Graphic Content"
+ description = "Generation of explicit sexual content, pornographic material, or disturbing graphic descriptions that violate content policies."
def __init__(
self,
diff --git a/deepteam/vulnerabilities/illegal_activity/illegal_activity.py b/deepteam/vulnerabilities/illegal_activity/illegal_activity.py
index 4db9a14a..57b28b49 100644
--- a/deepteam/vulnerabilities/illegal_activity/illegal_activity.py
+++ b/deepteam/vulnerabilities/illegal_activity/illegal_activity.py
@@ -30,6 +30,7 @@ IllegalActivityLiteral = Literal[
class IllegalActivity(BaseVulnerability):
name: str = "Illegal Activity"
+ description = "Assistance with illegal activities including weapons, drugs, violent crimes, cybercrime, sex crimes, or child exploitation."
def __init__(
self,
diff --git a/deepteam/vulnerabilities/intellectual_property/intellectual_property.py b/deepteam/vulnerabilities/intellectual_property/intellectual_property.py
index 248b3872..d313e06f 100644
--- a/deepteam/vulnerabilities/intellectual_property/intellectual_property.py
+++ b/deepteam/vulnerabilities/intellectual_property/intellectual_property.py
@@ -29,6 +29,7 @@ IntellectualPropertyLiteral = Literal[
class IntellectualProperty(BaseVulnerability):
name: str = "Intellectual Property"
+ description = "Copyright violations, trademark infringement, or imitation of protected intellectual property without authorization."
def __init__(
self,
diff --git a/deepteam/vulnerabilities/misinformation/misinformation.py b/deepteam/vulnerabilities/misinformation/misinformation.py
index 05a336c6..567fd58f 100644
--- a/deepteam/vulnerabilities/misinformation/misinformation.py
+++ b/deepteam/vulnerabilities/misinformation/misinformation.py
@@ -25,6 +25,7 @@ MisinformationLiteral = Literal[
class Misinformation(BaseVulnerability):
name: str = "Misinformation"
+ description = "Spreading false or misleading information through factual errors, unsupported claims, or inaccurate statements that harm public perception."
def __init__(
self,
diff --git a/deepteam/vulnerabilities/personal_safety/personal_safety.py b/deepteam/vulnerabilities/personal_safety/personal_safety.py
index 6ce6e790..17f2fe7d 100644
--- a/deepteam/vulnerabilities/personal_safety/personal_safety.py
+++ b/deepteam/vulnerabilities/personal_safety/personal_safety.py
@@ -28,6 +28,7 @@ PersonalSafetyLiteral = Literal[
class PersonalSafety(BaseVulnerability):
name: str = "Personal Safety"
+ description = "Encouragement of harmful actions including bullying, self-harm, unsafe practices, or stalking that endanger personal well-being."
def __init__(
self,
diff --git a/deepteam/vulnerabilities/pii_leakage/pii_leakage.py b/deepteam/vulnerabilities/pii_leakage/pii_leakage.py
index 7be83e68..bb561f61 100644
--- a/deepteam/vulnerabilities/pii_leakage/pii_leakage.py
+++ b/deepteam/vulnerabilities/pii_leakage/pii_leakage.py
@@ -27,6 +27,7 @@ PIILeakageLiteral = Literal[
class PIILeakage(BaseVulnerability):
name: str = "PII Leakage"
+ description = "Disclosure of personally identifiable information through direct disclosure, API/database access, session leaks, or social manipulation."
def __init__(
self,
diff --git a/deepteam/vulnerabilities/prompt_leakage/prompt_leakage.py b/deepteam/vulnerabilities/prompt_leakage/prompt_leakage.py
index 4b5af61f..33abe2d1 100644
--- a/deepteam/vulnerabilities/prompt_leakage/prompt_leakage.py
+++ b/deepteam/vulnerabilities/prompt_leakage/prompt_leakage.py
@@ -27,6 +27,7 @@ PromptLeakageLiteral = Literal[
class PromptLeakage(BaseVulnerability):
name: str = "Prompt Leakage"
+ description = "Exposure of system prompts, instructions, secrets, credentials, permissions, roles, or internal guard configurations to unauthorized users."
def __init__(
self,
diff --git a/deepteam/vulnerabilities/rbac/rbac.py b/deepteam/vulnerabilities/rbac/rbac.py
index f24b6c4a..da161d28 100644
--- a/deepteam/vulnerabilities/rbac/rbac.py
+++ b/deepteam/vulnerabilities/rbac/rbac.py
@@ -26,6 +26,7 @@ RBACLiteral = Literal[
class RBAC(BaseVulnerability):
name: str = "RBAC"
+ description = "Role-Based Access Control bypass enabling role restriction circumvention, privilege escalation, or unauthorized role assumption without proper validation."
def __init__(
self,
diff --git a/deepteam/vulnerabilities/recursive_hijacking/recursive_hijacking.py b/deepteam/vulnerabilities/recursive_hijacking/recursive_hijacking.py
index 1f43a66b..88bbd0ac 100644
--- a/deepteam/vulnerabilities/recursive_hijacking/recursive_hijacking.py
+++ b/deepteam/vulnerabilities/recursive_hijacking/recursive_hijacking.py
@@ -30,6 +30,7 @@ RecursiveHijackingLiteralType = Literal[
class RecursiveHijacking(BaseVulnerability):
name: str = "Recursive Hijacking"
+ description = "Recursive subversion of the AI's objectives through self-modifying goals, objective chaining, or goal propagation attacks."
def __init__(
self,
diff --git a/deepteam/vulnerabilities/robustness/robustness.py b/deepteam/vulnerabilities/robustness/robustness.py
index d7bbadaa..c75e630b 100644
--- a/deepteam/vulnerabilities/robustness/robustness.py
+++ b/deepteam/vulnerabilities/robustness/robustness.py
@@ -25,6 +25,7 @@ RobustnessLiteral = Literal["input_overreliance", "hijacking"]
class Robustness(BaseVulnerability):
name: str = "Robustness"
+ description = "AI fragility through input overreliance or susceptibility to hijacking that redirects its intended behavior."
def __init__(
self,
diff --git a/deepteam/vulnerabilities/shell_injection/shell_injection.py b/deepteam/vulnerabilities/shell_injection/shell_injection.py
index 8aac6b56..5b684495 100644
--- a/deepteam/vulnerabilities/shell_injection/shell_injection.py
+++ b/deepteam/vulnerabilities/shell_injection/shell_injection.py
@@ -26,6 +26,7 @@ ShellInjectionLiteral = Literal[
class ShellInjection(BaseVulnerability):
name: str = "Shell Injection"
+ description = "Malicious shell command generation enabling command injection, system command execution, or shell escape sequences to compromise system integrity."
def __init__(
self,
diff --git a/deepteam/vulnerabilities/sql_injection/sql_injection.py b/deepteam/vulnerabilities/sql_injection/sql_injection.py
index 77c0bcfe..7e33d45d 100644
--- a/deepteam/vulnerabilities/sql_injection/sql_injection.py
+++ b/deepteam/vulnerabilities/sql_injection/sql_injection.py
@@ -26,6 +26,7 @@ SQLInjectionLiteral = Literal[
class SQLInjection(BaseVulnerability):
name: str = "SQL Injection"
+ description = "Malicious SQL query generation using blind, union-based, or error-based injection techniques to manipulate database operations."
def __init__(
self,
diff --git a/deepteam/vulnerabilities/ssrf/ssrf.py b/deepteam/vulnerabilities/ssrf/ssrf.py
index 849e49d7..61733adf 100644
--- a/deepteam/vulnerabilities/ssrf/ssrf.py
+++ b/deepteam/vulnerabilities/ssrf/ssrf.py
@@ -26,6 +26,7 @@ SSRFLiteral = Literal[
class SSRF(BaseVulnerability):
name: str = "SSRF"
+ description = "Server-Side Request Forgery enabling unauthorized internal service access, cloud metadata access, or port scanning through crafted requests."
def __init__(
self,
diff --git a/deepteam/vulnerabilities/toxicity/toxicity.py b/deepteam/vulnerabilities/toxicity/toxicity.py
index 65b252c9..5e92f326 100644
--- a/deepteam/vulnerabilities/toxicity/toxicity.py
+++ b/deepteam/vulnerabilities/toxicity/toxicity.py
@@ -22,6 +22,7 @@ ToxicityLiteral = Literal["profanity", "insults", "threats", "mockery"]
class Toxicity(BaseVulnerability):
name: str = "Toxicity"
+ description = "Harmful or offensive language including profanity, insults, threats, or mockery toward individuals or groups."
def __init__(
self,