Agent identity & access

Every agent gets an identity.Every grant expires.

Each agent gets a cryptographic identity and scoped, expiring grants instead of standing credentials.

The problem

Agents inherit
more access than any person.

Most agents run on a shared service account or a long-lived key that reaches far more than any task needs.

  • Standing credentialsKeys that never expire, shared by several agents.
  • Over-scoped by defaultAn agent that answers questions can also write, export and email.
  • No way to tell agents apartLogs show the service account, not which agent acted, or for whom.
A real pathA copilot is asked to send a record outside.
  1. 1copilot-hr-assistThe agent starts and is attested: its identity, its owner and its declared task.
  2. 2grantIt receives a 15-minute grant to read the HR record of the employee in the ticket.
  3. 3documentA document in the ticket asks it to email that record to an outside address.
  4. Blockemail.send is outside the grant: the domain is not approved.copilot-hr-assist · email.send
Industry references
  • ASI03Identity and Privilege AbuseOWASP Top 10 for Agentic Apps
  • LLM06:2025Excessive AgencyOWASP Top 10 for LLMs
  • AML.T0055Unsecured CredentialsMITRE ATLAS
How Zenstra handles it

Least privilege,
per agent and per call.

Identity and access are decided by the same policy engine as every other action.

Identify

An identity for every agent

Each agent is attested when it starts, so every action traces back to one agent and the task it acted for.

  • Cryptographic identity per agent
  • Attestation checked before actions run
  • Revoked the moment access drifts
Scope

Grants that fit the task

Every tool an agent uses becomes a scoped grant: which collection, which member, which queue, and for how long.

  • Per-call and per-session grants
  • Short-lived credentials for Bedrock, Azure OpenAI and Vertex AI
  • Approved destinations for outbound actions
Revoke

One switch across every platform

Stop one agent everywhere it runs with a single revocation, instead of a switch per vendor.

  • Revocation in seconds
  • Grants listed for your risk register
  • Every grant and revocation on the record
Policy as code

The policy
behind it.

No standing credentials: each grant names one agent, one scope and an expiry, and anything outside it is refused.

  • Decided inline
  • Fails closed
  • Versioned with every decision
deny*when identity is not attestedallowhris.readfor employee in task · expires 15mdenyemail.sendwhen domain not in approved_domainsholdshell.execin prod · step-up required
Where it applies

Every connection becomes
a scoped grant.

  • Core and lending

    Connectors into core, LOS and CRM become scoped, expiring grants.

  • Model providers

    Per-agent credentials for AWS Bedrock, Azure OpenAI and Vertex AI.

  • MCP tools

    Each MCP tool call checked against the agent’s grant.

  • Memory and knowledge

    Signed writes only, so shared knowledge cannot be poisoned.

  • Outbound actions

    Email and exports limited to approved destinations.

  • Production access

    Shell and deploy actions held for a person to approve.

FAQ

Common questions.

Do agents need new credentials?

Zenstra issues short-lived credentials to each agent in place of standing keys. Your existing identity provider stays the source of truth for people.

What happens when an agent asks for more than its grant?

The action is blocked or held for a person to approve, according to your policy, and the decision is recorded with the reason.

Can we revoke one agent without breaking others?

Yes. Each agent has its own identity and grants, so revoking one leaves the rest running.