Skip to content

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

owasp updated

A-Vamshi566db6fe5ada
diff --git a/deepteam/frameworks/__init__.py b/deepteam/frameworks/__init__.py
index 06aef4b2..115a3933 100644
--- a/deepteam/frameworks/__init__.py
+++ b/deepteam/frameworks/__init__.py
@@ -1,5 +1,6 @@
-from .frameworks import OWASPTop10, AISafetyFramework
+from .frameworks import AISafetyFramework
from .aegis.aegis import Aegis
+from .owasp.owasp import OWASPTop10
__all__ = [
"AISafetyFramework",
diff --git a/deepteam/frameworks/aegis/aegis.py b/deepteam/frameworks/aegis/aegis.py
index 496e6524..82a121e9 100644
--- a/deepteam/frameworks/aegis/aegis.py
+++ b/deepteam/frameworks/aegis/aegis.py
@@ -1,8 +1,7 @@
import asyncio
import random
-from typing import Optional, Union, List
+from typing import Optional, Union
from deepeval.models import DeepEvalBaseLLM
-from deepeval.utils import get_or_create_event_loop
from tqdm import tqdm
from deepteam.frameworks import AISafetyFramework
diff --git a/deepteam/frameworks/frameworks.py b/deepteam/frameworks/frameworks.py
index 2e297517..04beed14 100644
--- a/deepteam/frameworks/frameworks.py
+++ b/deepteam/frameworks/frameworks.py
@@ -3,10 +3,6 @@ from typing import List, Optional
from deepeval.models import DeepEvalBaseLLM
-from deepteam.frameworks.utils import (
- _get_owasp_attacks,
- _get_owasp_vulnerabilities,
-)
from deepteam.vulnerabilities import (
BaseVulnerability,
)
@@ -22,24 +18,5 @@ class AISafetyFramework:
attacks: Optional[List[BaseAttack]]
_has_dataset: bool = False
- def load_dataset(self):
- pass
-
- def assess(self):
- pass
-
- async def a_assess(self):
- pass
-
class Config:
arbitrary_types_allowed = True
-
-
-class OWASPTop10(AISafetyFramework):
- def __init__(self):
- super().__init__(
- name="OWASP",
- description="The OWASP Top 10 for LLMs 2025",
- vulnerabilities=_get_owasp_vulnerabilities(),
- attacks=_get_owasp_attacks(),
- )
diff --git a/deepteam/frameworks/owasp/owasp.py b/deepteam/frameworks/owasp/owasp.py
new file mode 100644
index 00000000..fa2c34b0
--- /dev/null
+++ b/deepteam/frameworks/owasp/owasp.py
@@ -0,0 +1,212 @@
+from typing import List
+from deepteam.frameworks import AISafetyFramework
+from deepteam.vulnerabilities import (
+ BaseVulnerability,
+ Bias,
+ Toxicity,
+ Misinformation,
+ IllegalActivity,
+ PromptLeakage,
+ PIILeakage,
+ ExcessiveAgency,
+ Robustness,
+ IntellectualProperty,
+ Competition,
+ GraphicContent,
+ PersonalSafety,
+ RBAC,
+ BOLA,
+ BFLA,
+ SSRF,
+ DebugAccess,
+ ShellInjection,
+ SQLInjection,
+)
+
+from deepteam.attacks import BaseAttack
+from deepteam.attacks.single_turn import (
+ Base64,
+ GrayBox,
+ Leetspeak,
+ MathProblem,
+ Multilingual,
+ PromptInjection,
+ PromptProbing,
+ Roleplay,
+ ROT13,
+)
+
+from deepteam.attacks.multi_turn import (
+ CrescendoJailbreaking,
+ LinearJailbreaking,
+ TreeJailbreaking,
+ SequentialJailbreak,
+ BadLikertJudge,
+)
+
+class OWASPTop10(AISafetyFramework):
+
+ def __init__(self):
+ self.name = "OWASP"
+ self.description="The OWASP Top 10 for LLMs 2025"
+ self.vulnerabilities = self._get_owasp_vulnerabilities()
+ self.attacks = self._get_owasp_attacks()
+
+ def _get_owasp_vulnerabilities(self) -> List[BaseVulnerability]:
+ """
+ Get vulnerabilities based on OWASP Top 10 for LLMs 2025.
+
+ Reference: docs/red-teaming-owasp-top-10-for-llms.mdx
+ """
+ return [
+ # 1. Prompt Injection - Handled by attacks, not vulnerabilities
+ # 2. Sensitive Information Disclosure
+ PIILeakage(
+ types=[
+ "direct_disclosure",
+ "api_and_database_access",
+ "session_leak",
+ "social_manipulation",
+ ]
+ ),
+ PromptLeakage(
+ types=[
+ "secrets_and_credentials",
+ "instructions",
+ "guard_exposure",
+ "permissions_and_roles",
+ ]
+ ),
+ IntellectualProperty(
+ types=[
+ "imitation",
+ "copyright_violations",
+ "trademark_infringement",
+ "patent_disclosure",
+ ]
+ ),
+ # 3. Supply Chain - Not directly testable via attacks
+ # 4. Data and Model Poisoning
+ Bias(types=["race", "gender", "religion", "politics"]),
+ Toxicity(types=["profanity", "insults", "threats", "mockery"]),
+ Misinformation(
+ types=[
+ "factual_errors",
+ "unsupported_claims",
+ "expertize_misrepresentation",
+ ]
+ ),
+ IllegalActivity(
+ types=[
+ "cybercrime",
+ "violent_crimes",
+ "non_violent_crimes",
+ "illegal_drugs",
+ ]
+ ),
+ GraphicContent(
+ types=["sexual_content", "graphic_content", "pornographic_content"]
+ ),
+ PersonalSafety(
+ types=[
+ "bullying",
+ "self_harm",
+ "unsafe_practices",
+ "dangerous_challenges",
+ ]
+ ),
+ # 5. Improper Output Handling - Not directly testable via attacks
+ # 6. Excessive Agency
+ ExcessiveAgency(types=["functionality", "permissions", "autonomy"]),
+ # 7. System Prompt Leakage (covered under Sensitive Information Disclosure)
+ # 8. Vector and Embedding Weaknesses - Not directly testable via attacks
+ # 9. Misinformation (already covered above)
+ Competition(types=["discreditation"]),
+ # 10. Unbounded Consumption - Not directly testable via attacks
+ # Additional OWASP-related vulnerabilities
+ RBAC(
+ types=[
+ "role_bypass",
+ "privilege_escalation",
+ "unauthorized_role_assumption",
+ ]
+ ),
+ DebugAccess(
+ types=[
+ "debug_mode_bypass",
+ "development_endpoint_access",
+ "administrative_interface_exposure",
+ ]
+ ),
+ ShellInjection(
+ types=[
+ "command_injection",
+ "system_command_execution",
+ "shell_escape_sequences",
+ ]
+ ),
+ SQLInjection(
+ types=[
+ "blind_sql_injection",
+ "union_based_injection",
+ "error_based_injection",
+ ]
+ ),
+ BFLA(
+ types=[
+ "privilege_escalation",
+ "function_bypass",
+ "authorization_bypass",
+ ]
+ ),
+ BOLA(
+ types=[
+ "object_access_bypass",
+ "cross_customer_access",
+ "unauthorized_object_manipulation",
+ ]
+ ),
+ SSRF(
+ types=[
+ "internal_service_access",
+ "cloud_metadata_access",
+ "port_scanning",
+ ]
+ ),
+ Robustness(types=["hijacking", "input_overreliance"]),
+ ]
+
+
+ def _get_owasp_attacks(self) -> List[BaseAttack]:
+ """
+ Get attacks relevant for OWASP Top 10 for LLMs testing.
+ """
+ return [
+ # Core injection attacks
+ PromptInjection(weight=3),
+ # Encoding-based attacks for bypassing filters
+ Base64(weight=2),
+ ROT13(weight=2),
+ Leetspeak(weight=2),
+ # Social engineering attacks
+ Roleplay(
+ weight=2,
+ role="learned gentleman of natural philosophy",
+ persona="1600s Shakespearean scholar",
+ ),
+ PromptProbing(weight=2),
+ # Multilingual and obfuscation attacks
+ Multilingual(weight=2),
+ MathProblem(weight=1),
+ # Gray box attacks for known vulnerabilities
+ GrayBox(weight=2),
+ # Multi-turn sophisticated attacks
+ CrescendoJailbreaking(weight=2),
+ LinearJailbreaking(weight=2),
+ TreeJailbreaking(weight=1),
+ SequentialJailbreak(weight=2),
+ BadLikertJudge(weight=1),
+ ]
+
+ def get_name(self):
+ return "OWASPTop10"