Reference

    What Is AI Agent Security? Definition, Threats, Architecture & Controls

    Last updated August 2026Reviewed by the Andeavour research team4 cited sources

    In short

    AI agent security is the protection of an individual AI agent across its lifecycle: how it is created, what identity and credentials it holds, which tools it may call, what it remembers, what it is allowed to do, and how it is monitored and shut down. Its goal is to keep an autonomous agent inside an authorized, observable and reversible envelope of action.

    01

    Definition

    AI agent security is the set of controls applied to a single autonomous agent so that its actions stay authorized, attributable and reversible. Where agentic AI security describes the whole discipline, AI agent security is the per-agent implementation: identity, entitlements, tool scope, memory hygiene, runtime tracing and decommissioning.

    An AI agent is not an application in the classical sense. It has no fixed control flow: given a goal, the model selects tools and sequences them at runtime. Security therefore shifts from reviewing code paths to bounding capability — deciding in advance which actions the agent may take, under which conditions, with which data, and who is accountable for it.

    Because agents authenticate as non-human identities, agent security sits at the intersection of identity governance and application security. NIST's zero-trust guidance applies directly: authenticate every request, authorize per action, assume compromise and minimize implicit trust zones.

    02

    Security risks

    Standing credentials

    Agents often hold long-lived API keys or tokens so automation does not break. Those credentials outlive the task, the project and sometimes the team that created them.

    Unbounded tool scope

    An agent given a generic HTTP or shell tool effectively has unlimited capability, regardless of how narrow its stated purpose is.

    Untrusted input treated as instruction

    Retrieved documents, API responses and user messages all enter the same context window. Without trust separation, data becomes command.

    Persistent memory drift

    Agents that write to their own memory can accumulate manipulated or simply wrong facts that quietly steer future decisions.

    No ownership or offboarding

    When the creating team moves on, the agent keeps running with valid access and no accountable owner.

    Untraceable actions

    If only API calls are logged, an investigator can see what happened but never why the agent decided to do it.

    03

    Examples

    Finance reconciliation agent

    An agent granted write access to the ledger to fix mismatches. A crafted invoice line convinces it to approve a payment adjustment that no human reviewed.

    DevOps triage agent

    An incident agent with cluster admin restarts services during outages. An injected log line prompts it to scale down a production namespace.

    Research agent with browsing

    A market-research agent visits an attacker-controlled site that instructs it to fetch and post internal context to an external endpoint.

    Inherited orphan agent

    An agent built during a pilot keeps polling a data warehouse a year later; nobody remembers it exists until an audit finds the token.

    04

    Architecture

    Agent identity

    One credential per agent, short-lived, scoped to the data domain, mapped to a human owner and an expiry date.

    Policy decision point

    A runtime authority that evaluates every proposed tool call against policy before it executes, rather than trusting the model to self-restrict.

    Tool broker

    A mediated layer between the agent and real APIs that enforces allow-lists, validates parameters, redacts sensitive fields and rate-limits.

    Memory store with write policy

    Explicit rules on what may be written to long-term memory, with provenance on every entry and the ability to expire or quarantine.

    Trace pipeline

    Structured records of prompt, plan, tool call, arguments, result and policy decision, retained for investigation and replay.

    05

    Controls

    1

    Issue per-agent, short-lived credentials

    Replace shared service accounts and static keys with workload identity that expires automatically and is revocable in one action.

    2

    Bind tools to purpose

    Enumerate the exact operations the agent needs. Ban generic shell, generic HTTP and wildcard database access.

    3

    Enforce authorization outside the model

    Never rely on the system prompt to prevent an action. Policy must be evaluated by code the model cannot influence.

    4

    Quarantine untrusted content

    Tag retrieved content as data, strip instruction patterns, and prevent it from expanding scope or triggering privileged tools.

    5

    Require approval for irreversible actions

    Payments, deletions, permission grants and external communications should need a human or a second control to proceed.

    6

    Assign an owner and an expiry

    Every agent gets a named owner, a business justification and a review date, enforced by automation rather than a spreadsheet.

    06

    Comparison

    AspectHuman userService accountAI agent
    Decides next actionYes, with judgmentNo, fixed codeYes, at runtime from a model
    Susceptible to social engineeringYesNoYes, via prompt injection
    ScalesSlowlyPredictablyInstantly, in populations
    OffboardingHR-drivenChange-managedFrequently absent
    Explains its intentYesN/AOnly if traces are captured
    07

    Frequently asked questions

    Is an AI agent just a non-human identity?

    It has a non-human identity, but the identity is only one attribute. Unlike a static service account, an agent chooses its own actions, which is why intent-level controls and tracing matter.

    Can I secure agents with API gateways alone?

    A gateway is a good enforcement point for tool calls, but it does not know why the agent made the call, what content influenced it, or which Principal created the agent.

    How many agents do enterprises run?

    Most organizations cannot answer this, which is the point. Counts from copilots, developer tooling and SaaS automations typically exceed the documented inventory by a wide margin.

    What is the single highest-value control?

    Reducing standing privilege. Short-lived, narrowly scoped credentials cut the blast radius of every other failure mode.

    08

    Sources

    1. [1]Agentic AI — Threats and MitigationsOWASP GenAI Security Project
    2. [2]SP 800-207: Zero Trust ArchitectureNIST
    3. [3]What is agentic AI security?Microsoft Security
    4. [4]ATLAS — Adversarial Threat Landscape for AI SystemsMITRE
    09

    How Andeavour addresses it

    Per-agent identity governance

    AgentGuard inventories agents, maps each to an owner and enforces least privilege on the credentials they use.

    Action-level containment

    High-impact tool calls are gated, traced and revocable, so a manipulated agent cannot convert access into irreversible damage.

    Creation-layer coverage

    Because AgentGuard also secures AI Principals, newly spawned agents inherit policy instead of appearing outside governance.

    See it against your own environment

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