Skip to content

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

.

penguine17b4ffb37851
diff --git a/deepteam/frameworks/constants.py b/deepteam/frameworks/constants.py
index 07a6f927..639df40c 100644
--- a/deepteam/frameworks/constants.py
+++ b/deepteam/frameworks/constants.py
@@ -1,11 +1,15 @@
+from typing import Dict
from .aegis.aegis import Aegis
+from .frameworks import AISafetyFramework
from .beavertails.beavertails import BeaverTails
from .owasp.owasp import OWASPTop10
from .nist.nist import NIST
from .mitre.mitre import MITRE
-FRAMEWORKS_MAP = {f.name: f for f in [OWASPTop10, NIST, MITRE]}
+FRAMEWORKS_MAP: Dict[str, AISafetyFramework] = {
+ f.name: f for f in [OWASPTop10, NIST, MITRE]
+}
FRAMEWORK_RISK_CATEGORIES_MAP = {
f.name: f.ALLOWED_TYPES for f in [OWASPTop10, NIST, MITRE]
diff --git a/deepteam/frameworks/frameworks.py b/deepteam/frameworks/frameworks.py
index 0bf68459..e9774cb0 100644
--- a/deepteam/frameworks/frameworks.py
+++ b/deepteam/frameworks/frameworks.py
@@ -5,6 +5,7 @@ from deepteam.vulnerabilities import (
BaseVulnerability,
)
from deepteam.attacks import BaseAttack
+from deepteam.frameworks.risk_category import RiskCategory
@dataclass
@@ -13,6 +14,7 @@ class AISafetyFramework:
description: str
vulnerabilities: Optional[List[BaseVulnerability]]
attacks: Optional[List[BaseAttack]]
+ risk_categories: Optional[List[RiskCategory]]
_has_dataset: bool = False
class Config:
diff --git a/deepteam/frameworks/owasp/risk_categories.py b/deepteam/frameworks/owasp/risk_categories.py
index 2ddb5381..e62bfd70 100644
--- a/deepteam/frameworks/owasp/risk_categories.py
+++ b/deepteam/frameworks/owasp/risk_categories.py
@@ -22,7 +22,6 @@ from deepteam.vulnerabilities import (
SQLInjection,
CustomVulnerability,
)
-from deepteam.attacks import BaseAttack
from deepteam.attacks.single_turn import (
Base64,
GrayBox,
diff --git a/kk.py b/kk.py
deleted file mode 100644
index 7fb9ff83..00000000
--- a/kk.py
+++ /dev/null
@@ -1,24 +0,0 @@
-import random
-from deepteam import red_team
-from deepteam.frameworks import MITRE
-
-
-def mock_model_callback(prompt: str, turns=None) -> str:
- return "This request violates internal policy."
-
-
-categories = [
- "reconnaissance",
- "resource_development",
- "initial_access",
- "ml_attack_staging",
- "exfiltration",
- "impact",
-]
-random_category = random.choice(categories)
-risk_assessment = red_team(
- model_callback=mock_model_callback,
- framework=MITRE(categories=[random_category]),
- async_mode=False,
- ignore_errors=True,
-)