Topic primer

AI Agent Security 101

Why AI agents are not just another service account, the agent model, the threats that matter, and the principle that the model requests while authorization decides.

Trustivan SecurityUpdated August 202611 min read

Why agents are not just another service account

A service account has a fixed code path. Whatever it does tomorrow, it did yesterday, and you can read the code to find out. An AI agent assembles its behaviour at runtime from a goal, a model, the tools it can reach and the text it happens to read. The same agent, with the same credentials, can do something today that it has never done before — because a document told it to.

That difference changes four things about how agents must be governed:

  • Authority is inherited, not designed. Agents are typically given a developer’s credentials, a broad OAuth grant or a catch-all role, because scoping them properly requires knowing in advance what they will do — which is exactly what nobody knows.
  • Capabilities arrive at runtime. Tools and MCP servers can be discovered and called without anyone reviewing them. An agent’s capability set is a moving target.
  • Input is untrusted by construction. The model follows instructions in its context, and its context includes web pages, tickets, emails and tool results. Anyone who can get text in front of the agent can try to give it orders.
  • Actions can be irreversible and fast. Deleting a bucket, merging a pull request, sending a payment or messaging a customer happens at machine speed, often with nobody watching.

None of this means agents are ungovernable. It means the unit of governance is not the credential and not the prompt. It is the action: this identity, on this resource, for this purpose, in this context, with this authority.

The agent model

An agent is an identity with four extra edges

Every agent is a non-human identity — it has an owner, credentials, services it authenticates to and resources it can reach. What makes it an agent are four additional relationships that a service account does not have.

Uses tools

Functions, APIs, MCP servers and shells that turn model output into effects in the world

Delegates

Hands tasks — and authority — to other agents, which may hold more access than the delegator

Receives data

Reads documents, pages, tickets and tool results, any of which can carry instructions

Performs actions

Reads, writes, deploys, deletes and approves — some of them irreversible

Modelling agents this way has a practical consequence: an agent finding inherits everything the identity graph already knows. When an agent is flagged, the owner, the credentials, the effective access and the blast radius are already attached — and the agent-specific questions (which tools, on whose behalf, from which data) sit alongside them rather than in a separate inventory. Shadow AI, in particular, stops being a tooling problem and becomes a classification inside the identity program you already run.

The threats that matter

The agent threat vocabulary is large and growing. The ten below cover most real incidents. For each, the useful question is the same: whichidentity relationship does the attack exploit, and what control on that relationship would have stopped it?

ThreatHow it worksIdentity-centric control
Prompt injectionInstructions in the agent’s input redirect it toward the attacker’s goalAuthorize actions outside the model; the injected text can make the model request anything, not get anything
Indirect prompt injectionThe instructions arrive through content the agent reads — a page, a ticket, a tool resultTreat data sources as part of the threat surface; scope authority so that reading untrusted data cannot unlock high-impact actions
Tool poisoning and impersonationA tool is replaced, modified or spoofed so the agent calls something other than it believesTool identity and provenance; a tool the platform never discovered cannot be entitled, so the answer is DENY
Confused deputyThe agent uses its own broad authority to do something a less-privileged requester could notEvaluate the on-behalf-of identity and purpose, not only the agent’s permissions
Delegation abuseAn agent gains authority by delegating to, or being delegated by, an agent with more accessEvaluate effective authority across the whole delegation chain
Excessive autonomyThe agent takes irreversible, high-impact actions without a human in the loopCHALLENGE as a first-class verdict, answered by a grant somebody other than the requester approved
Memory poisoningPersistent memory or context is seeded with false facts or standing instructionsProvenance for what the agent was given; decisions that do not trust the agent’s own state
Cross-tenant and cross-session leakageData or instructions from one tenant or session reach another through shared agents or toolsTenant-scoped identity and credentials per agent; no shared standing access
Denial of walletThe agent is driven to burn paid resources — tokens, calls, compute — at harmful scalePer-identity budgets and rate limits as part of action policy; anomalous consumption as a risk signal
Credential replay and chainingCredentials held in the agent’s context are reused or combined beyond their intended scopeBroker short-lived, action-scoped credentials instead of handing the agent standing ones

OWASP maintains a broader catalogue of agentic threats; we map its categories to identity-centric controls on adedicated page.

The authorization principle

The model can request an action. The authorization system decides.Available

A language model is never the security boundary, because the attacker and the defender share the same channel: text. The boundary has to be made of things the attacker’s text cannot reach — identity, policy and evidence evaluated outside the model. The flow below is an illustrative action; the gate, its rules and its verdicts are built, no connector discovers an agent, and the tool entitlements it decides over are declared by an operator.

Illustrative action on a sample environment. Two details matter: CHALLENGE is a verdict rather than an error, and the decision names the rules that produced it in a record a database trigger refuses to edit.

Every high-impact action should pass through the same evaluation, regardless of which model asked for it:

Identity
Who is acting — resolved from the credential presented, never claimed in the request body.
Credential
Whether the credential is live, unexpired and not revoked.
Tool
Whether an operator entitled this tool to this agent. Nothing entitled, no write goes through; reads are not gated.
Effect
Whether the call reads or mutates.
Resource
The impact an operator assigned it. An unclassified resource is not treated as harmless.
Autonomy
Whether the agent was declared supervised or unsupervised.
Grant
Whether an access grant, approved by somebody other than the requester, is standing behind the action.
Decision
ALLOW · CHALLENGE · DENY — strictest verdict wins, with the rules that fired recorded.

A stated purpose is deliberately not on that list. The gate records the intent a caller declares, but no rule decides on it: a purpose is exactly what an attacker writes freely, so it is kept for the investigator reading the refusal rather than used to justify it.

Most actions are low impact and should be allowed instantly with evidence recorded. The point is that the few that matter are evaluated by something that cannot be talked out of it. The research paper develops the argument in full.

How Trustivan approaches it

Trustivan’s approach to AI agent security is to govern, secure and deploy agents responsibly across the enterprise — on top of the identity graph that already exists for every other non-human identity. Labels below are deliberate: agent governance and the runtime gate are available, no connector discovers an agent, and no screen registers one — the gate decides over tool entitlements an operator declares.

  • The identity graph agents plug into

    Available

    Owners, identities, credentials, services and resources with evidence on every edge. No connector discovers an agent, and no screen registers one; agents reach a deployment today only through the synthetic demo seeder.

    Identity Graph
  • Agent discovery & inventory

    Not built

    No connector discovers agents, copilots or orchestration identities, by design. MCP servers and the tools they list are discovered by the MCP connector.

    Agent Discovery
  • Agent identity & ownership

    Available

    An owner, a declared autonomy and a lifecycle for every agent in the inventory, with the same ownership and quarantine flow as any other non-human identity.

    Agent Identity
  • Tools & MCP

    Available

    MCP servers and the tools they list, with the effect a publisher declares or an operator sets, so an agent’s tools are inventoried and an unclassified tool is visible as one.

    MCP Security
  • Agent permissions

    Available

    Effective access of an agent identity on the graph, as an upper bound, and the tools an operator entitled it to. What an agent actually calls is never observed.

    Agent Permissions
  • Provenance & shadow AI

    Not built

    No agent lineage edge and no endpoint telemetry, by design. Agent lineage reports as unavailable rather than guessed, and an agent identity without a profile raises shadow_agent_unprofiled.

    Agent Provenance
  • Runtime authorization

    Available

    A prompt-injection-resistant gate between the model and the tool. The acting principal comes from the presented credential, the tool entitlements it decides over are declared by an operator, and it enforces when the agent runtime calls it.

    Agent Runtime Security
  • Action policy & decisions

    Available

    Thirteen deny-by-default rules over credential state, tool entitlement, effect, resource impact, declared autonomy and standing grants, returning ALLOW, CHALLENGE or DENY with the rule that produced it recorded.

    Secure Every Action
  • Multi-agent trust

    Not built

    Delegation chains and agent-to-agent trust are not evaluated, by design: no connector reports an edge between agents, so each action is decided on the presenting agent’s own credential.

    Multi-Agent Security

What we will not claim. Trustivan is not an AI firewall, an MCP gateway or an agent observability product. Those tools see prompts, traffic and traces. The control plane’s job is the relationship between an agent’s identity, its authority and its actions — which is what makes a decision defensible.

See an agent action evaluated end to end.

A demo walks through an agent, its tools, a high-impact action and the rules behind a CHALLENGE or DENY verdict — on sample data.

Govern every agent. Secure every action.

See how TRUSTIVAN connects identity, credential, access, agent and action context into one control plane.