Skip to main content
Oleria gives each AI agent a risk level so you can fix the riskiest agents first. The level comes from the agent’s risk reasons: specific conditions such as an agent anyone can use, or an agent that shares its account with other agents. This page explains each level and each reason, why it matters, and where to check it.
Agent discovery is enabled per workspace. To turn it on, contact us at support@oleria.com.

Risk reasons at a glance

Oleria has 17 risk reasons. The table lists them most severe first, the order in which the Risk tooltip lists an agent’s reasons. Select a reason for what it means, why it matters, and everything to check.

Risk levels

An agent’s risk level is the highest severity among its risk reasons. Oleria evaluates an agent’s risk each time the agent’s own integration syncs, from what Oleria holds at that moment about every connected application. A change made in another connected application shows only after that application’s integration has synced and the agent’s integration has synced after it.

Where risk appears

How to read risk reasons

Some risk reasons combine two simpler ones. When a combined reason applies, Oleria lists it in place of the reasons it combines: The Risk Finding filter matches the reasons an agent lists, so filtering on a simple reason doesn’t return an agent where a combined reason replaced it. To find every agent with a simple reason, select it together with every combined reason whose row names it. For Runs as an account with elevated privilege, that is four combined reasons. The first two combined reasons also apply to an agent that can act as others and has no elevated-privilege account, so the filter returns those agents too. Each reason below shows its on-screen title and severity. The “Why it matters” and “What to check” notes are Oleria’s guidance, not text you see in the product. Two terms recur in the reasons below:
  • An elevated-privilege account is an account the agent runs as whose Privilege is Elevated on the Agent Identities tab.
  • An agent can act as others when an account it runs as, or a group or role that account belongs to, can act as an account, group, or role other than the agent’s own accounts.

Who can use the agent

Anyone can use it without signing in

Severity: High The agent’s audience is Public, and a grant lets anyone use it without signing in. Why it matters: Anyone who can reach the agent can make it act, and there is no identity to attribute an interaction to, limit, or revoke. What to check: On the Who Can Use tab, find the principal that stands for anyone, such as an account whose Type is Anonymous user, and the entry point under Via. If the agent isn’t meant to be public, remove that grant in the source application.

Set to allow anyone, but not reachable yet

Severity: Low The agent’s audience is Public, but no grant lets an anonymous caller use it. Why it matters: The agent is configured to be public. One grant to an anonymous principal opens it to anyone. What to check: Confirm that the Audience of Public is intended. If it isn’t, change the agent’s audience in the source application.

Principals outside the organization can use it

Severity: Moderate Someone outside your organization can use the agent, and Anyone can use it without signing in doesn’t apply. This includes a grant to anyone on an agent whose audience isn’t Public. Why it matters: People you don’t manage can make the agent act with the access its accounts hold. What to check: On the Who Can Use tab, look for principals whose Type is External user, for groups and roles that reach beyond your organization, and for a principal that stands for anyone, such as a group named for anonymous or all users. On AWS, a resource policy that allows any principal shows this way. Confirm that each one is intended.

Anyone can use it, and it runs as an account with elevated privilege or can act as others

Severity: Critical Anyone can use the agent without signing in, and the agent either runs as an elevated-privilege account or can act as others. Why it matters: An anonymous caller can steer an agent that holds elevated-privilege access or can borrow another identity’s access. What to check: Find the public grant on the Who Can Use tab. On the Agent Identities tab, look for an account whose Privilege is Elevated. When no account is Elevated, the reason comes from what the agent’s accounts can act as:
  • Select an account’s name on the Agent Identities tab to open the account’s own page. For a non-human identity (NHI), look on the Access tab there for access the account gets by acting as another account or a group. When it acts as another account, Impersonation names that account.
  • On the Access Graph, with Access selected under Graph types, select the account. When its side panel has a Can impersonate list, look there for accounts it can act as directly.
  • Neither place shows a right the account holds through a group or role it belongs to, or a right to act as a role. When neither place explains the reason, check in the source application which accounts, groups, and roles the account, and each group and role it belongs to, can impersonate or assume.
Close the public access first, then reduce what the agent’s accounts can reach and act as.

Outside principals can use it, and it runs as an account with elevated privilege or can act as others

Severity: Critical Someone outside your organization can use the agent, and the agent either runs as an elevated-privilege account or can act as others. Why it matters: People you don’t manage can steer an agent that holds elevated-privilege access or can borrow another identity’s access. What to check: Find the outside principals on the Who Can Use tab. Then check the agent’s accounts as described under Anyone can use it, and it runs as an account with elevated privilege or can act as others: first for a Privilege of Elevated, then for what they can act as. Remove the outside access that isn’t intended, then reduce what the agent’s accounts can reach and act as.

Anyone can use it, and its tools run on a credential every caller uses

Severity: High Anyone can use the agent without signing in, and something the agent depends on runs on a stored credential instead of the caller’s own. Why it matters: Every anonymous caller acts with that credential’s access. What to check: On the Composition tab, find the rows that can cause the reason, as described under Its tools run on a credential every caller uses. On the Who Can Use tab, find the public grant. Close the public access, or change the dependency to use each caller’s own authority.

The accounts it runs as

Runs as an account with elevated privilege

Severity: Low An account the agent runs as has a Privilege of Elevated. Why it matters: Whatever the agent does, or is manipulated into doing, runs with that account’s access. What to check: Find the account on the Agent Identities tab, then see what it can reach on the Access tab. Reduce the account’s access to what the agent needs.

Shares an account with other agents

Severity: Moderate Another agent runs as one of this agent’s accounts. Why it matters: Activity on a shared account can’t be tied to one agent, and each agent holds the access the others need. A shared account doesn’t count toward the agent’s Last Activity. What to check: On the Agent Identities tab, Shared With names the other agents. Give each agent its own account in the source application.

A credential on its accounts is risky

Severity: Moderate A credential that authenticates one of the agent’s accounts has a risk of its own. Why it matters: Whoever obtains that credential can act as the agent’s account. What to check: On the Access Graph, a Credential node connects to the account it authenticates. See NHI secret risk recommendations for each NHI secret risk reason and its fix.

Someone can become one of its accounts

Severity: Moderate A principal other than the agent can act as one of the accounts the agent runs as. Why it matters: That principal gets the account’s access, and what it does through the account looks like the agent’s activity. What to check: Open the agent on the Access Graph with Access selected under Graph types. The Can impersonate this agent’s accounts list names each principal and the account it can act as. Remove the right where it isn’t needed. The list also shows principals the source platform itself operates. Those don’t cause this reason.

Someone can become one of its accounts with elevated privilege

Severity: Critical A principal other than the agent can act as one of the agent’s accounts that has elevated privilege. Why it matters: That principal gets elevated-privilege access, and what it does through the account looks like the agent’s activity. What to check: Use the Can impersonate this agent’s accounts list as described under Someone can become one of its accounts. The Agent Identities tab shows which account has a Privilege of Elevated.

How the agent acts

Its tools run on a credential every caller uses

Severity: Moderate Something the agent depends on runs on a stored credential instead of the caller’s own, and other principals can use the agent. For an agent with no account of its own, the reason applies only when the source platform doesn’t check each caller’s own right to that credential. Why it matters: Each caller acts with the stored credential’s access, whatever their own access is. What to check: On the Composition tab, a row can cause this reason only when the source application names a credential for it, its Runs With isn’t Delegation or Impersonation, and its State isn’t Disabled. Start with the rows that show a Credential. A row whose Credential cell shows a dash can cause it too. The dash appears when the source application names a credential that Oleria doesn’t hold, which counts, and when it names none, which doesn’t. To tell those rows apart, check the dependency’s credential in the source application. On the Who Can Use tab, see who can use the agent. Limit who can use the agent, or change the dependency to use each caller’s own authority.

Acts autonomously with tools that run without approval

Severity: Low The agent’s autonomy is Autonomous, and a tool, Model Context Protocol (MCP) server, or delegate it depends on is declared to run with no approval step. Why it matters: The agent can use those dependencies with no person involved. What to check: On the Composition tab, find the rows whose Approval is Automatic and whose Used As is Tool, MCP server, or Delegate. Decide which of them should need a person’s approval.

Runs as an account with elevated privilege and acts autonomously without approvals

Severity: Moderate The agent runs as an elevated-privilege account, its autonomy is Autonomous, and a tool, MCP server, or delegate is declared to run with no approval step. Why it matters: The agent can use elevated-privilege access with no person involved. What to check: Check the account on the Agent Identities tab and the Approval column on the Composition tab. Reduce the account’s access, or require approval for the dependencies that act on it.

No guardrail is configured

Severity: Low A live version of the agent, or the agent itself when it has no versions, has dependencies and no guardrail among them. Why it matters: Nothing recorded can block what the agent receives, produces, or does. What to check: Open the Controls tab. Oleria sees only the guardrails the source application reports, so confirm in the source application before you act.

Its live configuration can change in place

Severity: Low A live version of the agent can change without becoming a new version. Why it matters: The agent’s behavior can change with no new version to review or return to. What to check: On the Versions tab, find the versions whose Immutable value is No.

Ownership

No one is accountable for it

Severity: Moderate The agent has no accountable owner. Why it matters: No one answers for what the agent does, reviews its access, or decides when to retire it. What to check: See Owners and Created By on the agent details page. Record an owner for the agent in the source application. If Owners lists a group and the reason still applies, Oleria hasn’t counted the group’s members yet. See Owners.

Contact us

For questions, contact us at support@oleria.com.