arxiv_ai 90% Match Survey Paper AI Security Researchers,Cybersecurity Professionals,AI Developers,Policy Makers 1 week ago

Agentic AI Security: Threats, Defenses, Evaluation, and Open Challenges

ai-safety › alignment

📄 Abstract

Abstract: Agentic AI systems powered by large language models (LLMs) and endowed with planning, tool use, memory, and autonomy, are emerging as powerful, flexible platforms for automation. Their ability to autonomously execute tasks across web, software, and physical environments creates new and amplified security risks, distinct from both traditional AI safety and conventional software security. This survey outlines a taxonomy of threats specific to agentic AI, reviews recent benchmarks and evaluation methodologies, and discusses defense strategies from both technical and governance perspectives. We synthesize current research and highlight open challenges, aiming to support the development of secure-by-design agent systems.

Authors (4)

Shrestha Datta

Shahriar Kabir Nahin

Anshuman Chhabra

Prasant Mohapatra

Submitted

October 27, 2025

arXiv Category

cs.AI

arXiv PDF

Key Contributions

This survey provides a comprehensive overview of the security landscape for agentic AI systems powered by LLMs. It introduces a taxonomy of novel threats, reviews existing evaluation methodologies, and discusses defense strategies, highlighting the unique security risks posed by autonomous AI agents and outlining open research challenges.

Business Value

Crucial for organizations developing or deploying agentic AI systems, enabling them to proactively identify and mitigate security risks, protect sensitive data, and ensure the safe operation of autonomous agents.

Paper Metadata

Innovation Type

Survey / Taxonomy

Deployment Feasibility

High relevance for deployment, as it informs security best practices and risk management for agentic AI systems.

Limitations Addressed

Addresses the emerging and amplified security risks associated with agentic AI systems, which are distinct from traditional AI safety and software security concerns. It aims to provide a structured understanding of these new threats and defenses.

Performance Gains

N/A (Survey paper)

Technical Tags

agentic AIsecurity threatsAI safetyLLM securityautonomytool usebenchmarksdefense strategies

Research Topics

AI SafetyCybersecurityAgent SystemsLarge Language ModelsAI Ethics

Methods & Architectures

Taxonomy of threatsReview of benchmarks and evaluation methodologiesDiscussion of defense strategiesSynthesis of current research Agentic AI systemsLarge Language Models (LLMs)

Applications & Tasks

Cybersecurity AI Development Autonomous Systems SecurityRisk AssessmentThreat ModelingDefense Identifying and categorizing security threats in agentic AIReviewing evaluation methods for agentic AI securityDiscussing defense strategies and open challenges

Datasets & Benchmarks

Benchmarks

Recent benchmarks for agentic AI security (unspecified)

Security metrics (unspecified)

Related Fields

CybersecurityArtificial IntelligenceAI SafetySoftware SecurityAutonomous Systems

Keywords

agentic AIsecurityLLMthreatsdefensesAI safetyautonomytool usebenchmarkscybersecurityautonomous agentsrisk assessmentAI governance

Academic Context

#AI Safety#Cybersecurity#Agent Systems#Large Language Models#AI Ethics

Commercial Potential

Potential Products

AI security assessment toolsSecure agent development frameworks

Target Industries

TechnologyFinanceDefenseAny industry deploying AI agents

Use Case Examples

Assessing the security vulnerabilities of AI agents used in financial tradingDeveloping secure protocols for autonomous drone operationsProtecting AI agents from adversarial attacks

Competitive Edge

Provides a foundational understanding and structured framework for the nascent field of agentic AI security, guiding future research and development.

Resource Requirements

Compute Needs

N/A (Survey paper)

Data Requirements

N/A (Survey paper)

Deployment Constraints

Understanding and mitigating the unique attack vectors specific to autonomous, tool-using AI agents.

Scalability

The principles discussed are relevant for scaling agentic AI systems securely.

View Full Paper Back to Papers