Skip to content

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

new owasp 2026 framework added

A-Vamshi32d11dbfef1a
diff --git a/deepteam/frameworks/constants.py b/deepteam/frameworks/constants.py
index bc9c240f..c6968ef7 100644
--- a/deepteam/frameworks/constants.py
+++ b/deepteam/frameworks/constants.py
@@ -7,13 +7,15 @@ from .owasp.owasp import OWASPTop10
from .nist.nist import NIST
from .mitre.mitre import MITRE
from .risk_category import RiskCategory
+from .owasp_top_10_agentic.owasp_top_10_agentic import OWASP_ASI_2026
from .owasp.risk_categories import OWASP_CATEGORIES
from .nist.risk_categories import NIST_CATEGORIES
from .mitre.risk_categories import MITRE_CATEGORIES
+from .owasp_top_10_agentic.risk_categories import OWASP_ASI_CATEGORIES
FRAMEWORKS_MAP: Dict[str, AISafetyFramework] = {
- f.name: f for f in [OWASPTop10, NIST, MITRE]
+ f.name: f for f in [OWASPTop10, NIST, MITRE, OWASP_ASI_2026]
}
DATASET_FRAMEWORKS_MAP = {f.name: f for f in [Aegis, BeaverTails]}
@@ -71,4 +73,10 @@ FRAMEWORK_RISK_CATEGORY_MAPPING: Dict[str, FrameworkInfo] = {
_serialize_risk_category(rc) for rc in MITRE_CATEGORIES
],
),
+ OWASP_ASI_2026.name: FrameworkInfo(
+ description=OWASP_ASI_2026.description,
+ risk_categories=[
+ _serialize_risk_category(rc) for rc in OWASP_ASI_CATEGORIES
+ ]
+ )
}
diff --git a/deepteam/frameworks/owasp_top_10_agentic/owasp_top_10_agentic.py b/deepteam/frameworks/owasp_top_10_agentic/owasp_top_10_agentic.py
new file mode 100644
index 00000000..0ed7c805
--- /dev/null
+++ b/deepteam/frameworks/owasp_top_10_agentic/owasp_top_10_agentic.py
@@ -0,0 +1,100 @@
+from typing import List, Literal
+from deepteam.frameworks import AISafetyFramework
+from deepteam.frameworks.owasp_top_10_agentic.risk_categories import OWASP_ASI_CATEGORIES
+
+"""
+OWASP Top 10 for Agentic Applications (ASI) 2026
+===============================================
+
+The OWASP Top 10 for Agentic Applications identifies the most critical security risks
+introduced by autonomous and semi-autonomous AI agents. Unlike traditional LLM
+applications, agentic systems combine reasoning, memory, tools, and multi-step execution,
+introducing new classes of vulnerabilities that extend beyond prompt-level attacks.
+
+The 2026 edition focuses on failures arising from goal misalignment, tool misuse,
+delegated trust, inter-agent communication, persistent memory, and emergent autonomous
+behavior.
+
+The 10 critical risks are:
+1. ASI01 - Agent Goal Hijack: Manipulation of agent objectives or plans
+2. ASI02 - Tool Misuse & Exploitation: Unsafe or excessive use of tools
+3. ASI03 - Agent Identity & Privilege Abuse: Misuse of delegated authority or trust
+4. ASI04 - Agentic Supply Chain Compromise: Malicious tools, agents, or metadata
+5. ASI05 - Unexpected Code Execution: Unsafe execution of agent-generated code
+6. ASI06 - Memory & Context Poisoning: Corruption of agent memory or contextual state
+7. ASI07 - Insecure Inter-Agent Communication: Manipulated agent-to-agent messages
+8. ASI08 - Cascading Agent Failures: Failure propagation across agent systems
+9. ASI09 - Human-Agent Trust Exploitation: Abuse of human over-reliance on agents
+10. ASI10 - Rogue Agents: Agents acting beyond intended objectives
+
+Each category includes:
+- Attacks: Prompting and interaction patterns that induce unsafe agent behavior
+- Vulnerabilities: Typed behavioral weaknesses observable at the model or agent layer
+
+This framework evaluates agentic risks at the reasoning and behavioral level and does not
+claim full coverage of runtime enforcement, infrastructure security, or cryptographic
+protocols.
+
+Reference: https://genai.owasp.org/resource/owasp-top-10-for-agentic-applications-for-2026/
+"""
+
+
+class OWASP_ASI_2026(AISafetyFramework):
+ name = "OWASP Top 10 for Agentic Applications 2026"
+ description = "A comprehensive list of the most critical security risks associated with agentic AI applications. The 2026 edition focuses on failures introduced by autonomy, tool usage, delegated trust, memory persistence, inter-agent communication, and emergent behavior. Each risk category is evaluated using realistic attack techniques and typed vulnerability assessments aligned with agent reasoning and behavior."
+
+ ALLOWED_TYPES = [
+ "ASI_01",
+ "ASI_02",
+ "ASI_03",
+ "ASI_04",
+ "ASI_05",
+ "ASI_06",
+ "ASI_07",
+ "ASI_08",
+ "ASI_09",
+ "ASI_10",
+ ]
+
+ def __init__(
+ self,
+ categories: List[
+ Literal[
+ "ASI_01",
+ "ASI_02",
+ "ASI_03",
+ "ASI_04",
+ "ASI_05",
+ "ASI_06",
+ "ASI_07",
+ "ASI_08",
+ "ASI_09",
+ "ASI_10",
+ ]
+ ] = [
+ "ASI_01",
+ "ASI_02",
+ "ASI_03",
+ "ASI_04",
+ "ASI_05",
+ "ASI_06",
+ "ASI_07",
+ "ASI_08",
+ "ASI_09",
+ "ASI_10",
+ ],
+ ):
+ self.categories = categories
+ self.risk_categories = []
+ self.vulnerabilities = []
+ self.attacks = []
+
+ for category in categories:
+ for risk_category in OWASP_ASI_CATEGORIES:
+ if risk_category.name == category:
+ self.risk_categories.append(risk_category)
+ self.vulnerabilities.extend(risk_category.vulnerabilities)
+ self.attacks.extend(risk_category.attacks)
+
+ def get_name(self):
+ return self.name
diff --git a/deepteam/frameworks/owasp_top_10_agentic/risk_categories.py b/deepteam/frameworks/owasp_top_10_agentic/risk_categories.py
new file mode 100644
index 00000000..6d60c79b
--- /dev/null
+++ b/deepteam/frameworks/owasp_top_10_agentic/risk_categories.py
@@ -0,0 +1,186 @@
+from typing import List
+from deepteam.frameworks.risk_category import RiskCategory
+from deepteam.vulnerabilities import (
+ Ethics,
+ Misinformation,
+ PromptLeakage,
+ BFLA,
+ BOLA,
+ RBAC,
+ DebugAccess,
+ ShellInjection,
+ SQLInjection,
+ IntellectualProperty,
+ GoalTheft,
+ RecursiveHijacking,
+ Robustness,
+ ExcessiveAgency,
+ IndirectInstruction,
+ ToolOrchestrationAbuse,
+ AgentIdentityAbuse,
+ ToolMetadataPoisoning,
+ UnexpectedCodeExecution,
+ InsecureInterAgentCommunication,
+ AutonomousAgentDrift,
+)
+from deepteam.attacks.single_turn import (
+ Base64,
+ PromptInjection,
+ PromptProbing,
+ Roleplay,
+ ROT13,
+ ContextPoisoning,
+)
+from deepteam.attacks.multi_turn import (
+ CrescendoJailbreaking,
+ LinearJailbreaking,
+ TreeJailbreaking,
+)
+
+
+
+
+OWASP_ASI_CATEGORIES: List[RiskCategory] = [
+ RiskCategory(
+ name="ASI_01",
+ _display_name="ASI01:2026 Agent Goal Hijack",
+ description="Attackers manipulate agent goals, plans, or decision paths through direct or indirect instruction injection, causing agents to pursue unintended or malicious objectives.",
+ attacks=[
+ PromptInjection(weight=3),
+ LinearJailbreaking(weight=2),
+ CrescendoJailbreaking(weight=2),
+ TreeJailbreaking(weight=1),
+ Roleplay(weight=2, role="System planner", persona="Authorized orchestrator"),
+ ],
+ vulnerabilities=[
+ GoalTheft(types=["escalating_probing", "cooperative_dialogue"]),
+ ExcessiveAgency(types=["autonomy", "functionality"]),
+ Robustness(types=["hijacking"]),
+ RecursiveHijacking(types=["goal_propagation_attacks", "self_modifying_goals"]),
+ IndirectInstruction(types=["cross_context_injection", "document_embedded_instructions", "rag_injection", "tool_output_injection"])
+ ],
+ ),
+ RiskCategory(
+ name="ASI_02",
+ _display_name="ASI02:2026 Tool Misuse & Exploitation",
+ description="Agents misuse or abuse tools through unsafe composition, recursion, or excessive execution, causing harmful side effects despite valid permissions.",
+ attacks=[
+ PromptInjection(weight=2),
+ Roleplay(weight=2, role="Automation agent", persona="Agent orchestrator"),
+ ],
+ vulnerabilities=[
+ ExcessiveAgency(types=["functionality", "permissions"]),
+ BFLA(types=["function_bypass", "authorization_bypass"]),
+ ToolOrchestrationAbuse(types=["recursive_tool_calls", "unsafe_tool_composition", "tool_budget_exhaustion", "cross_tool_state_leakage"]),
+ ],
+ ),
+ RiskCategory(
+ name="ASI_03",
+ _display_name="ASI03:2026 Agent Identity & Privilege Abuse",
+ description="Abuse of delegated authority, ambiguous agent identity, or trust assumptions leading to unauthorized actions.",
+ attacks=[
+ Roleplay(weight=3, role="Privileged agent", persona="Authorized priority agent"),
+ PromptProbing(weight=2),
+ ],
+ vulnerabilities=[
+ BOLA(types=["object_access_bypass", "unauthorized_object_manipulation"]),
+ RBAC(types=["role_bypass", "unauthorized_role_assumption"]),
+ PromptLeakage(types=["guard_exposure", "permissions_and_roles", "secrets_and_credentials"]),
+ AgentIdentityAbuse(types=["agent_impersonation", "cross_agent_trust_abuse", "identity_inheritance"])
+ ],
+ ),
+ RiskCategory(
+ name="ASI_04",
+ _display_name="ASI04:2026 Agentic Supply Chain Compromise",
+ description="Compromise of external agents, tools, schemas, or prompts that agents dynamically trust or import.",
+ attacks=[
+ PromptInjection(weight=2),
+ ],
+ vulnerabilities=[
+ IntellectualProperty(types=["imitation"]),
+ ToolMetadataPoisoning(types=["schema_manipulation","description_deception","permission_misrepresentation","registry_poisoning"]),
+ ],
+ ),
+ RiskCategory(
+ name="ASI_05",
+ _display_name="ASI05:2026 Unexpected Code Execution",
+ description="Agent-generated or agent-triggered code execution without sufficient validation or isolation.",
+ attacks=[
+ PromptInjection(weight=2),
+ Base64(weight=2),
+ ROT13(weight=2),
+ ],
+ vulnerabilities=[
+ ShellInjection(types=["command_injection", "system_command_execution"]),
+ SQLInjection(types=["blind_sql_injection"]),
+ DebugAccess(types=["administrative_interface_exposure", "debug_mode_bypass"]),
+ UnexpectedCodeExecution(types=["unauthorized_code_execution","shell_command_execution","eval_usage"]),
+ ],
+ ),
+ RiskCategory(
+ name="ASI_06",
+ _display_name="ASI06:2026 Memory & Context Poisoning",
+ description="Injection or leakage of agent memory or contextual state that influences future reasoning or actions.",
+ attacks=[
+ PromptInjection(weight=2),
+ CrescendoJailbreaking(weight=1),
+ ContextPoisoning(weight=3)
+ ],
+ vulnerabilities=[
+ Misinformation(types=["expertize_misrepresentation", "factual_errors"]),
+ Robustness(types=["input_overreliance", "hijacking"]),
+ ],
+ ),
+ RiskCategory(
+ name="ASI_07",
+ _display_name="ASI07:2026 Insecure Inter-Agent Communication",
+ description="Manipulation of messages exchanged between agents, planners, and executors.",
+ attacks=[
+ Roleplay(weight=2, role="Planner", persona="Critical Thinking Agent"),
+ LinearJailbreaking(weight=2),
+ CrescendoJailbreaking(weight=1)
+ ],
+ vulnerabilities=[
+ InsecureInterAgentCommunication(types=["agent_in_the_middle", "message_injection", "message_spoofing"]),
+ ],
+ ),
+ RiskCategory(
+ name="ASI_08",
+ _display_name="ASI08:2026 Cascading Agent Failures",
+ description="Small agent failures propagate through connected systems causing large-scale impact.",
+ attacks=[
+ PromptInjection(weight=2),
+ LinearJailbreaking(weight=1),
+ CrescendoJailbreaking(weight=1),
+ ],
+ vulnerabilities=[
+ Robustness(types=["hijacking"]),
+ ExcessiveAgency(types=["autonomy"]),
+ ],
+ ),
+ RiskCategory(
+ name="ASI_09",
+ _display_name="ASI09:2026 Human-Agent Trust Exploitation",
+ description="Exploitation of human over-reliance on agents through misleading explanations or authority framing.",
+ attacks=[
+ Roleplay(weight=3, persona="Authoritative expert", role="Human Agent"),
+ ],
+ vulnerabilities=[
+ Ethics(types=["moral_integrity", "responsible_transparency"]),
+ Misinformation(types=["expertize_misrepresentation", "unsupported_claims"]),
+ ],
+ ),
+ RiskCategory(
+ name="ASI_10",
+ _display_name="ASI10:2026 Rogue Agents",
+ description="Agents acting beyond intended objectives due to goal drift, collusion, or emergent behavior.",
+ attacks=[
+ CrescendoJailbreaking(weight=1),
+ LinearJailbreaking(weight=2),
+ ],
+ vulnerabilities=[
+ RecursiveHijacking(types=["goal_propagation_attacks", "self_modifying_goals"]),
+ AutonomousAgentDrift(types=["agent_collusion","goal_drift","reward_hacking","runaway_autonomy"]),
+ ],
+ ),
+]
diff --git a/deepteam/metrics/__init__.py b/deepteam/metrics/__init__.py
index 7fa6ea33..5117f82a 100644
--- a/deepteam/metrics/__init__.py
+++ b/deepteam/metrics/__init__.py
@@ -24,7 +24,13 @@ from .hijacking.hijacking import HijackingMetric
from .intellectual_property.intellectual_property import (
IntellectualPropertyMetric,
)
-
+from .indirect_instruction.indirect_instruction import IndirectInstructionMetric
+from .tool_orchestration.tool_orchestration import ToolOrchestrationMetric
+from .agent_identity_abuse.agent_identity_abuse import AgentIdentityAbuseMetric
+from .tool_metadata_poisoning.tool_metadata_poisoning import ToolMetadataPoisoningMetric
+from .unexpected_code_execution.unexpected_code_execution import UnexpectedCodeExecutionMetric
+from .insecure_inter_agent_communication.insecure_inter_agent_communication import InsecureInterAgentCommunicationMetric
+from .autonomous_agent_drift.autonomous_agent_drift import AutonomousAgentDriftMetric
# Agentic metrics
from .agentic.unauthorized_execution.unauthorized_execution import (
diff --git a/deepteam/vulnerabilities/__init__.py b/deepteam/vulnerabilities/__init__.py
index 92e6ee47..226f297d 100644
--- a/deepteam/vulnerabilities/__init__.py
+++ b/deepteam/vulnerabilities/__init__.py
@@ -16,6 +16,13 @@ from .shell_injection.shell_injection import ShellInjection
from .sql_injection.sql_injection import SQLInjection
from .ssrf.ssrf import SSRF
from .intellectual_property.intellectual_property import IntellectualProperty
+from .indirect_instruction.indirect_instruction import IndirectInstruction
+from .tool_orchestration_abuse.tool_orchestration_abuse import ToolOrchestrationAbuse
+from .agent_identity_abuse.agent_identity_abuse import AgentIdentityAbuse
+from .tool_metadata_poisoning.tool_metadata_poisoning import ToolMetadataPoisoning
+from .unexpected_code_execution.unexpected_code_execution import UnexpectedCodeExecution
+from .insecure_inter_agent_communication.insecure_inter_agent_communication import InsecureInterAgentCommunication
+from .autonomous_agent_drift.autonomous_agent_drift import AutonomousAgentDrift
from .competition.competition import Competition
from .graphic_content.graphic_content import GraphicContent
from .personal_safety.personal_safety import PersonalSafety
@@ -45,6 +52,13 @@ __all__ = [
"SQLInjection",
"SSRF",
"IntellectualProperty",
+ "IndirectInstruction",
+ "ToolOrchestrationAbuse",
+ "AgentIdentityAbuse",
+ "ToolMetadataPoisoning",
+ "UnexpectedCodeExecution",
+ "InsecureInterAgentCommunication",
+ "AutonomousAgentDrift",
"Competition",
"GraphicContent",
"PersonalSafety",
diff --git a/deepteam/vulnerabilities/constants.py b/deepteam/vulnerabilities/constants.py
index 45c507bc..b1d84531 100644
--- a/deepteam/vulnerabilities/constants.py
+++ b/deepteam/vulnerabilities/constants.py
@@ -23,7 +23,13 @@ from .goal_theft.goal_theft import GoalTheft
from .recursive_hijacking.recursive_hijacking import RecursiveHijacking
from .robustness.robustness import Robustness
from .excessive_agency.excessive_agency import ExcessiveAgency
-
+from .indirect_instruction.indirect_instruction import IndirectInstruction
+from .tool_orchestration_abuse.tool_orchestration_abuse import ToolOrchestrationAbuse
+from .agent_identity_abuse.agent_identity_abuse import AgentIdentityAbuse
+from .tool_metadata_poisoning.tool_metadata_poisoning import ToolMetadataPoisoning
+from .unexpected_code_execution.unexpected_code_execution import UnexpectedCodeExecution
+from .insecure_inter_agent_communication.insecure_inter_agent_communication import InsecureInterAgentCommunication
+from .autonomous_agent_drift.autonomous_agent_drift import AutonomousAgentDrift
# Import types
from .bias.types import BiasType
@@ -47,6 +53,13 @@ from .goal_theft.types import GoalTheftType
from .recursive_hijacking.types import RecursiveHijackingType
from .robustness.types import RobustnessType
from .excessive_agency.types import ExcessiveAgencyType
+from .indirect_instruction.indirect_instruction import IndirectInstructionType
+from .tool_orchestration_abuse.tool_orchestration_abuse import ToolOrchestrationAbuseType
+from .agent_identity_abuse.agent_identity_abuse import AgentIdentityAbuseType
+from .tool_metadata_poisoning.tool_metadata_poisoning import ToolMetadataPoisoningType
+from .unexpected_code_execution.unexpected_code_execution import UnexpectedCodeExecutionType
+from .insecure_inter_agent_communication.insecure_inter_agent_communication import InsecureInterAgentCommunicationType
+from .autonomous_agent_drift.autonomous_agent_drift import AutonomousAgentDriftType
VULNERABILITY_CLASSES_MAP: Dict[str, BaseVulnerability] = {
@@ -73,6 +86,13 @@ VULNERABILITY_CLASSES_MAP: Dict[str, BaseVulnerability] = {
RecursiveHijacking,
Robustness,
ExcessiveAgency,
+ IndirectInstruction,
+ ToolOrchestrationAbuse,
+ AgentIdentityAbuse,
+ ToolMetadataPoisoning,
+ UnexpectedCodeExecution,
+ InsecureInterAgentCommunication,
+ AutonomousAgentDrift,
]
}