Get Early Access
HomeLearnAI Security

AI SECURITY

What is AI security?

AI security is the practice of protecting AI systems, applications, models, data, identities, tools, and actions from misuse, attack, unauthorized access, and unsafe behavior.

As AI systems evolve from generating answers to taking actions, security increasingly needs to address not only what AI says, but what it can access and do.

Explore AI Agent SecurityExplore AI Security Threats
AI SECURITYFrom model to action
MODELAPPLICATIONCONTEXT + DATAAGENTTOOLS + MCPIDENTITY + ACTION

THE DEFINITION

AI security protects the whole system

AI security is the practice of protecting artificial intelligence systems from attacks, misuse, unauthorized access, data exposure, and unsafe behavior. It includes securing models, applications, training and retrieval data, identities, tools, infrastructure, and the actions AI systems can perform.

  • Model and application security
  • Prompt injection defense
  • Data protection
  • Identity and access control
  • AI agent security
  • Tool and MCP security
  • Runtime monitoring and policy
  • AI red teaming
  • Supply-chain security
  • Governance and auditability

WHY IT EXISTS

Why does AI need its own security controls?

Instructions and data can appear in the same context.
Model behavior is probabilistic.
Untrusted content can influence decisions.
Models may reveal sensitive context.
AI systems may use tools and APIs.
Agents may make autonomous decisions.
AI can ingest many external sources.
Third-party models and components create dependencies.

AI security extends established security disciplines; it does not make IAM, AppSec, cloud, network, or data security obsolete.

CONCEPTUAL BOUNDARIES

How is AI security different from cybersecurity?

AI security sits within broader cybersecurity. It adds AI-specific attack surfaces and behaviors to existing security practices.

CYBERSECURITYProtects systems, networks, applications, users, and data from digital threats.IAM · AppSec · Cloud · Network · Data
AI SECURITYApplies those principles to models, prompts, context, agents, tools, retrieval data, and autonomous actions.Models · Agents · Tools · Runtime

MODELS AND SYSTEMS

LLM security is part of AI security, not the whole category

LLM SECURITYPrompt injection, sensitive disclosure, unsafe model output, and model misuse.
AI SECURITYLLMs, traditional ML, agents, model infrastructure, data, identities, tools, runtime actions, and supply chains.

LLM security focuses on a class of models. AI security follows the wider system around them.

THE ACTION BOUNDARY

AI security vs. AI agent security

AI security protects AI systems broadly. AI agent security focuses on systems that use tools, access applications, call APIs, operate with identities, retrieve data, make decisions, and execute actions.

What is AI Agent Security? Explore AI Agent Security
MODEL SECURITYWhat does the model produce?↓AGENT SECURITYWhat can the system do?Action creates a new security boundary.

THE ATTACK SURFACE

What does the AI attack surface include?

Vulnerabilities and unsafe trust relationships can exist at every layer of an AI system.

01USERS / INPUTS↓
02AI APPLICATION↓
03MODEL↓
04CONTEXT / RAG / MEMORY↓
05AGENT↓
06TOOLS / MCP / APIs↓
07IDENTITIES / CREDENTIALS↓
08ENTERPRISE SYSTEMS↓
09DATA / ACTIONS

RISK TAXONOMY

What are the main AI security risks?

Prompt InjectionMalicious instructions influence AI behavior.Indirect Prompt InjectionExternal content manipulates an AI system.Sensitive Data ExposureInformation is revealed or moved beyond its intended boundary.Excessive AgencyAgents receive too much autonomy.Excessive PermissionsAgents can access more than their task requires.Tool MisuseLegitimate tools are used in unsafe ways.MCP RiskMCP expands tool and data trust relationships.
Model / Supply Chain RiskThird-party models, libraries, plugins, and infrastructure add dependencies.
PoisoningTraining, retrieval, or persistent context may be manipulated.
Insecure Output HandlingGenerated output creates downstream risk when trusted without validation.

PROMPT INJECTION

Prompt injection is one of the defining AI application risks

AI systems process natural-language instructions and content. Attackers may exploit that boundary by embedding malicious or conflicting instructions.

Direct injection comes from an attacker interacting with the AI. Indirect injection sits inside external content the AI later retrieves. The risk becomes more consequential when the AI can use tools or take actions.

Prompt Injection Indirect Prompt Injection
UNTRUSTED CONTENT↓AI CONTEXT↓TOOL / DATA ACCESS↓REAL-WORLD ACTION

DATA SECURITY

AI security is also data security

Security teams need to understand what enters prompts and context, what the system can retrieve, what persists, where output goes, and who can access it.

Training dataPromptsUploaded filesRAG sourcesContext + memoryTool outputGenerated outputLogs + embeddingsCustomer information
Explore AI Agent Data Exfiltration

IDENTITY + ACCESS

AI systems need identity and access boundaries

Agents may authenticate through service accounts, workload identities, OAuth tokens, API keys, cloud roles, or delegated user access. Security teams need to understand which identity is used, what permissions it has, what systems it reaches, and whether that access fits the task.

AI security increasingly overlaps with non-human identity security.

Explore Agent Identity & Permissions
Service accountWorkload identityOAuth tokenAPI keyCloud roleDelegated user

RUNTIME SECURITY

Some AI security decisions must happen at runtime

Static permissions ask whether an identity can access a system. Runtime security asks whether an agent should perform an action now.

Agent identityRequested toolActionData sensitivityResourceEnvironmentAutonomyDestination
Explore Runtime Security
ALLOWExpected low-risk action
REVIEWSensitive action needs context
BLOCKAction violates policy

MCP SECURITY

MCP adds new tool and trust relationships

Model Context Protocol gives AI applications and agents a standard way to connect to external tools and data. Security questions include which servers exist, who operates them, which credentials sit behind them, and whether returned content can influence the agent.

What is MCP Security? Explore MCP Security MCP Attacks
AGENT↓MCP SERVER↓TOOL + CREDENTIAL↓DOWNSTREAM SYSTEM

ADVERSARIAL TESTING

AI systems should be tested adversarially

Red teaming tests how an AI system behaves when inputs, content, tools, permissions, or workflows are manipulated.

Prompt injectionIndirect injectionTool misuseExcessive accessApproval bypassMCP manipulationUnsafe multi-step workflowsData exfiltration
Explore AI Red Teaming What is AI Red Teaming?

SECURITY LIFECYCLE

AI security should span the system lifecycle

01DISCOVERKnow which AI systems and agents exist.
02ASSESSUnderstand models, data, identities, tools, and access.
03TESTFind weaknesses before deployment.
04CONTROLApply identity, permission, and runtime policy.
05MONITORObserve behavior and actions.
06INVESTIGATEUnderstand what happened.
07IMPROVEUpdate controls as systems change.

BEFORE + AFTER DEPLOYMENT

AI security happens before and after deployment

BUILD-TIME
  • Architecture review
  • Threat modeling
  • Permission design
  • Model evaluation
  • Red teaming
  • Tool review
RUNTIME
  • Action monitoring
  • Access enforcement
  • Sensitive-data controls
  • Approval workflows
  • Anomaly investigation
  • Audit trail

ENTERPRISE COVERAGE

What should enterprises secure?

AI APPLICATIONSInterfaces and workflows using models.
MODELSHosted, self-hosted, or third-party models.
DATAPrompts, RAG data, files, context, and memory.
AGENTSSystems that reason, use tools, and act.
IDENTITIESHuman and non-human access.
TOOLSFunctions, APIs, plugins, and MCP tools.
INFRASTRUCTURECloud, endpoints, pipelines, and hosting.
ACTIONSWrites, exports, transactions, messages, and execution.

PRACTICAL GUIDANCE

AI security best practices

01INVENTORY AI SYSTEMSKnow which models, applications, and agents exist.
02APPLY LEAST PRIVILEGELimit access to what each AI workload needs.
03TREAT EXTERNAL CONTENT AS UNTRUSTEDDo not assume retrieved text has instruction authority.
04PROTECT SENSITIVE DATAMinimize what enters prompts, context, and workflows.
05SECURE TOOLSUnderstand tool capabilities and permissions.
06CONTROL HIGH-IMPACT ACTIONSUse runtime policy and human review where appropriate.
07RED TEAM AI SYSTEMSTest adversarial scenarios before production.
08SECURE MCP CONNECTIONSUnderstand servers, tools, credentials, and downstream systems.
09MAINTAIN AUDITABILITYRecord identities, actions, decisions, and outcomes.
10REVIEW CONTINUOUSLYReassess as models, tools, data, and permissions evolve.

AUTHORITATIVE GUIDANCE

AI security frameworks and guidance

Use recognized frameworks as shared language for threat modeling and control design. They do not replace an organization’s own risk analysis.

OWASPGuidance covering risks in LLM and agentic AI applications.
MITRE ATLASA knowledge base for adversarial techniques targeting AI systems.
NISTAI risk-management and cybersecurity guidance.

OPERATING MODEL

Who is responsible for AI security?

AI security is cross-functional because AI systems cross traditional security boundaries.

CISO / SECURITYStrategy and risk.
APPSECAI-enabled application risks.
CLOUD SECURITYInfrastructure and permissions.
IDENTITY SECURITYHuman and non-human identities.
DATA SECURITYSensitive information and movement.
AI PLATFORMModels, agents, tools, and infrastructure.
ENGINEERINGApplication implementation.
GRC / RISKPolicies, controls, and governance.

FOR SECURITY TEAMS

Questions security teams should ask about AI

EYRANE’S POINT OF VIEW

AI security changes when AI can act

The security model changes when AI systems move from generating information to taking action.

Eyrane is focused on that boundary: discovering AI agents, understanding the identities and permissions behind them, and controlling sensitive actions at runtime.

DISCOVER→UNDERSTAND→CONTROL
Explore AI Agent Security

COMMON QUESTIONS

Questions about AI security

What is AI security?

AI security is the practice of protecting artificial intelligence systems from attacks, misuse, unauthorized access, data exposure, and unsafe behavior. It covers models, applications, data, identities, tools, infrastructure, and actions.

Why is AI security important?

AI security is important because AI systems can process sensitive information, consume untrusted content, connect to tools, and increasingly make decisions or take actions across enterprise systems.

How is AI security different from cybersecurity?

AI security applies cybersecurity principles to AI-specific systems and risks, including models, prompts, context, agents, tools, retrieval data, and autonomous actions. It complements IAM, AppSec, cloud, network, and data security.

What is generative AI security?

Generative AI security protects systems that create text, code, images, or other content from prompt injection, sensitive disclosure, misuse, unsafe outputs, and weaknesses in the surrounding application.

What is LLM security?

LLM security focuses on risks around large language models and the applications that use them. LLM security is part of AI security, which also includes agents, identities, tools, data, infrastructure, and runtime actions.

What is AI agent security?

AI agent security protects AI systems that use tools, access applications, call APIs, operate with identities, retrieve data, make decisions, and execute actions.

What are the main AI security risks?

Major risks include prompt injection, indirect prompt injection, sensitive data exposure, excessive agency, overprivileged agents, tool misuse, MCP risks, poisoning, supply-chain risk, and insecure output handling.

What is prompt injection?

Prompt injection is an attack that places malicious or conflicting instructions in an AI system’s input or context to influence behavior, reveal information, misuse tools, or cause unauthorized actions.

What is indirect prompt injection?

Indirect prompt injection hides malicious instructions in content an AI later retrieves, such as web pages, documents, email, tool output, RAG sources, or MCP responses.

How can companies secure AI agents?

Companies can inventory agents, map identities and permissions, minimize data, treat external content as untrusted, secure tools and MCP connections, apply runtime policy, red team systems, and maintain auditability.

What is AI runtime security?

AI runtime security evaluates AI behavior while a system is operating, using context such as identity, tool, action, resource, data sensitivity, environment, and autonomy to allow, review, or block sensitive operations.

What is MCP security?

MCP security protects the servers, tools, credentials, data, and downstream systems connected through Model Context Protocol, including trust relationships and actions taken through those tools.

What is AI red teaming?

AI red teaming tests AI systems with adversarial inputs and scenarios to find unsafe behavior, control failures, excessive access, tool misuse, injection paths, and dangerous multi-step actions.

What does least privilege mean for AI systems?

Least privilege means giving each AI workload only the data, identity permissions, tools, actions, duration, and environment it needs for its task.

Who is responsible for AI security?

AI security is cross-functional. CISO and security teams, AppSec, cloud and identity security, data security, AI platform teams, engineering, and GRC may all own parts of the system.

RELATED LEARNING

Continue learning about AI security

What is AI Agent Security?What is Agentic Security?What is MCP Security?What is AI Security Posture Management?What is AI Red Teaming?What is AI Agent Identity?OWASP Agentic Top 10MITRE ATLAS

RELATED THREATS

Explore AI security threats

Prompt InjectionIndirect Prompt InjectionSensitive Data ExposureExcessive AgencyExcessive PermissionsTool MisuseMCP Risk

AI SECURITY

Secure AI from model to action.

Protect AI agents across identities, permissions, tools, MCP connections, data, and runtime actions.

Explore AI Agent SecurityGet Early Access