Reference

    What Is Agentic AI Security? Definition, Risks, Architecture & Controls

    Last updated August 2026Reviewed by the Andeavour research team5 cited sources

    In short

    Agentic AI security is the practice of protecting autonomous AI agents — software that plans, decides and acts on its own — along with the tools, data and identities they use, and the AI Principals that create, train and command them. It covers agent discovery, identity and least privilege, prompt-injection defense, tool and memory integrity, runtime monitoring and blast-radius containment.

    01

    Definition

    Agentic AI security is the discipline of securing systems in which AI models do not just generate text but take actions: calling APIs, executing code, moving money, changing configuration and invoking other agents. Because the model chooses the next step at runtime, the security boundary is no longer a fixed application path — it is a decision loop that must be constrained, observed and revocable.

    The scope is wider than model safety. It spans the agent runtime (prompts, plans, tool calls, memory), the non-human identities agents authenticate with, the tool and connector surface such as Model Context Protocol servers, the data the agent reads and writes, and the creation layer: the programs, pipelines and orchestrators — AI Principals — that spawn agents, fine-tune their models and issue their instructions.

    Microsoft frames agentic AI security around governing autonomous agents with the same rigor applied to human identities, while OWASP's GenAI Security Project catalogues agent-specific threats such as memory poisoning, tool misuse, privilege compromise and cascading multi-agent failures. Both point at the same conclusion: agents need identity, authorization, observability and containment as first-class controls.

    02

    Security risks

    Prompt injection and instruction hijacking

    Untrusted content retrieved by an agent — a web page, ticket, PDF or email — can carry instructions the model follows. Because agents hold credentials and tools, an injection becomes an action, not just a bad answer.

    Excessive agency and over-privilege

    Agents are commonly granted broader access than a human performing the same task, because scoping is hard and speed wins. A single compromised agent then reaches far more data than its job required.

    Shadow agents

    Agents created inside SaaS platforms, copilots and developer tooling frequently never enter any inventory, so no one owns their credentials, scope or decommissioning.

    Memory and knowledge poisoning

    Long-lived agent memory, vector stores and fine-tuning data are writable attack surfaces. Poisoned entries persist across sessions and survive credential rotation.

    Tool and supply-chain abuse

    Connectors, MCP servers and plugins extend an agent's reach. A malicious or compromised tool definition can redirect the agent's actions without touching the model itself.

    Cascading multi-agent failure

    When agents call agents, one bad output becomes another agent's trusted input. Errors and injections propagate at machine speed with no human in the loop.

    03

    Examples

    Support agent exfiltrating data via a ticket

    A customer-service agent reads an incoming ticket containing hidden instructions telling it to summarize recent accounts and post the summary to an external webhook. The agent has legitimate CRM read access, so nothing looks anomalous at the identity layer.

    Coding agent with a standing cloud token

    A build agent holds long-lived cloud credentials so it can deploy. A poisoned dependency instructs it to add an IAM role during a routine change. The action is authorized, logged and wrong.

    Copilot-created shadow agents

    Business users assemble automations inside a productivity suite. Each one authenticates with delegated permissions, none appear in the CMDB, and offboarding never touches them.

    Poisoned retrieval corpus

    An attacker seeds an internal wiki page that an agent retrieves for policy questions. Every downstream agent that consults the corpus inherits the manipulated policy.

    04

    Architecture

    Creation layer (AI Principals)

    Programs, orchestrators and training pipelines that spawn agents, shape their models and issue their instructions. Compromise here mass-produces compromised agents.

    Identity layer

    Non-human identities, workload credentials, delegated OAuth grants and API keys. This is where least privilege, short-lived credentials and ownership are enforced.

    Reasoning layer

    The model, system prompt, planner and memory. Controls here include prompt hardening, trust separation between instructions and data, and memory write policies.

    Action layer

    Tools, connectors, MCP servers and APIs. Every capability should be individually allow-listed, parameter-validated and rate-limited.

    Observability and control plane

    Full traces of prompts, plans, tool calls and outcomes, with policy enforcement, human approval gates for high-impact actions, and an emergency kill switch.

    05

    Controls

    1

    Discover every agent and Principal

    Build a continuous inventory of agents, the identities they use and the programs that create them. You cannot govern a population you cannot count.

    2

    Give every agent a scoped, owned identity

    No shared service accounts. Short-lived credentials, explicit human ownership, automatic expiry and offboarding tied to the workflow that created the agent.

    3

    Separate instructions from data

    Treat all retrieved content as untrusted. Strip or neutralize instruction-like content, and never let retrieved text elevate the agent's tool scope.

    4

    Allow-list tools and validate parameters

    Bind each agent to the minimum tool set, validate arguments server-side, and require signed, versioned tool definitions.

    5

    Gate high-impact actions

    Require human approval or step-up verification for irreversible operations: payments, permission changes, deletions, production deploys and external sends.

    6

    Log the reasoning, not just the request

    Persist prompts, plans, tool calls, arguments and results so incidents can be reconstructed and replayed.

    7

    Contain blast radius

    Segment agents by data domain, cap spend and call volume, and keep a tested kill switch that revokes credentials for an agent population in one action.

    06

    Comparison

    Agentic AI security overlaps with adjacent disciplines but is not covered by any of them.

    DisciplinePrimary objectWhat it misses for agents
    Traditional AppSecCode and deployed applicationsNo concept of a runtime-chosen action path or model-issued decisions
    LLM securityModel inputs and outputsStops at the answer; ignores tools, credentials and persistent memory
    IAM / NHI governanceIdentities and entitlementsSees the credential, not the intent or the instruction that triggered its use
    EDR / XDREndpoints and processesTreats agent runtimes as generic applications behaving normally
    Agentic AI securityAgents, tools, memory and AI PrincipalsCombines identity, intent, action and creation-layer control
    07

    Frequently asked questions

    How is agentic AI security different from AI security?

    AI security is broad and includes model theft, training-data protection and safety. Agentic AI security narrows to systems that act autonomously, where the primary risk is unauthorized or manipulated action rather than an incorrect answer.

    Do existing IAM tools cover AI agents?

    Partly. IAM can issue and revoke credentials, but it does not see the prompt, plan or tool call that caused a credential to be used, and it has no notion of the AI Principal that created the agent.

    Is prompt filtering enough?

    No. Prompt filtering reduces some injection attempts but cannot constrain what an agent is permitted to do. Authorization, tool allow-listing and approval gates are what actually bound the damage.

    Where should an enterprise start?

    Start with discovery: inventory agents, their identities and their creators. Then scope privileges down, add tracing, and gate irreversible actions.

    08

    Sources

    1. [1]What is agentic AI security?Microsoft Security
    2. [2]Agentic AI — Threats and MitigationsOWASP GenAI Security Project
    3. [3]AI Risk Management Framework (AI RMF 1.0)NIST
    4. [4]ATLAS — Adversarial Threat Landscape for AI SystemsMITRE
    5. [5]Secure AI Framework (SAIF)Google
    09

    How Andeavour addresses it

    AgentGuard secures agents and Principals

    AgentGuard discovers agents and the AI Principals that create, train and command them, enforces least privilege on non-human identities, and contains blast radius when a Principal or agent is compromised.

    AgentHunter defends against external agentic attacks

    AgentHunter identifies autonomous adversarial agents at the edge and mitigates AI-driven reconnaissance and abuse campaigns in real time.

    Agentic SOC automates investigation and response

    The Agentic SOC Intelligence Agent correlates agent telemetry with existing security tooling so agent incidents are investigated and contained without adding analyst headcount.

    See it against your own environment

    A 30-minute technical session with the Andeavour team — no integration required to start.