AI agent identity security: accounts, permissions and approvals
20 October 2026
Improve AI agent identity security with practical checks for dedicated accounts, scoped access, credential rotation, revocation and human approval.
AI agent identity security needs attention before an automated workflow gets access to your business systems. In a CRN interview, 1Password CTO Nancy Wang raised concerns about applying human-oriented identity controls to agents that start, stop and act across multiple systems.
For Australian SMBs, the practical question is straightforward: can you identify each agent, limit what it can do and withdraw its access quickly? A private Australian-hosted environment with human approval retained provides the operating model. Explicit access controls make that model enforceable.
Why AI agent identity security needs separate controls
Wang’s comments highlight a lifecycle problem. An employee usually has a continuing account and an established role. An agent may run briefly, launch supporting tasks and return later under a new execution environment. Its credentials must not become unmanaged leftovers.
Existing identity principles still matter: least privilege, accountable ownership and access reviews. What changes is how consistently those controls must apply to short-lived workloads and repeated automated actions.
Separate three things in your design:
- Agent identity: which approved workflow is acting.
- Run identity: which particular execution produced an action.
- Human authority: who approved the workflow and any consequential changes.
A stable workflow identity can support traceability while individual runs use short-lived credentials. The right implementation depends on the connected applications, but every action should be attributable to a workflow and a specific run.
Give each agent its own accountable identity
Do not run production agents through an employee’s everyday login. That mixes human and automated activity, makes investigations harder and can leave the workflow dependent on someone’s employment or password changes.
Use an application identity, service principal or dedicated service account supported by the target system. Separate identities by workflow and environment rather than giving every agent one shared account.
Create an identity register containing:
- Business purpose and approved systems.
- Business owner and technical support owner.
- Production or test designation.
- Authentication method and credential location.
- Granted permissions and approval requirements.
- Review date and retirement procedure.
For example, an Order and Data Intake Agent should not inherit the identity used by a Financial and Management Reporting Agent. Their tasks, data needs and permitted actions differ.
Where a legacy system requires an interactive account, document that exception. Restrict its use and do not bypass MFA or share a staff browser session simply to make automation work.
Scope permissions to the task, not the department
Effective AI agent identity security starts with a list of required operations, not a broad job title such as “finance assistant”.
For each connected system, specify what the agent needs to read, create, update, send or delete. Narrow access to particular sites, libraries, records or endpoints wherever the platform supports it.
An AP/AR Automation Agent might read invoices and purchase orders, identify mismatches and prepare drafts. That does not justify permission to change supplier bank details, approve payments or release funds.
Check permissions at both layers:
- Identity layer: what the application registration, account or token permits.
- Application layer: which records and operations the destination system actually allows.
If an API only offers broad access, consider a restricted integration layer that exposes approved operations. Record the remaining risk rather than treating a prompt instruction as a security boundary.
Test denied actions as well as successful ones. Attempt an unauthorised deletion, unrelated record lookup and unapproved write. The system should block them independently of the model’s response.
Rotate credentials and control where secrets travel
Prefer short-lived tokens or workload identity federation where supported. These approaches reduce reliance on long-lived secrets, although issued access still needs monitoring and a workable revocation path.
Where passwords, API keys or client secrets are unavoidable:
- Store them in a managed secrets store, not prompts, documents or source code.
- Retrieve them through the integration layer without exposing them to the model where possible.
- Limit which workloads can retrieve each secret.
- Assign an owner, expiry date and rotation procedure.
- Exclude credentials from logs, exports and support attachments.
Credential rotation is not complete when a new secret is created. Update the integration, test the workflow and remove the old credential after a controlled transition. Check that cached sessions or forgotten copies do not preserve access.
Set rotation requirements according to credential type, exposure and system capability. Rotate immediately when compromise is suspected; a calendar schedule is not an incident response.
Make revocation a tested operational procedure
Stopping an agent process does not necessarily remove its access. Tokens, browser sessions, queued jobs and downstream integrations may remain active.
An AI agent identity security checklist needs a named person who can stop the workflow and a documented procedure to:
- Pause triggers, schedules and queued work.
- Disable the agent’s identity or remove its permissions.
- Revoke refresh tokens, sessions and secrets where supported.
- Block downstream access if existing tokens remain valid.
- Preserve logs and inspect recent actions.
- Verify that new requests and queued actions fail.
Some access tokens can remain usable until expiry, depending on the service. Document that limitation and the compensating controls available, rather than assuming account disablement is instantaneous everywhere.
Test revocation before production and after significant integration changes. Our OAuth token theft checklist explains why token access deserves attention beyond password resets.
Keep human approval outside the agent’s control
Human approval must be enforced by the workflow, not merely requested in its instructions. The agent should not be able to approve its own proposal, alter approval records or use an alternative route to execute a blocked action.
For payment-related changes, customer communications and production record updates, define:
- What the agent may prepare without approval.
- Which actions require an authorised reviewer.
- What evidence the reviewer must see.
- When an approval expires.
- Which changes require fresh approval.
Bind approval to the specific proposed action. If an amount, recipient or destination changes afterwards, require another review. Keep the approver’s identity distinct from the agent’s execution identity.
For a practical finance workflow, see AI agents for invoice processing: where humans approve.
Verify hosting boundaries and evidence before launch
Tech Engine’s managed agents run in a private Australian-hosted environment and retain human approval. However, hosting location alone does not define every data flow. Connected SaaS applications, external model endpoints, telemetry and backups also need assessment.
Map where documents, prompts, responses, credentials and audit records travel. Confirm which services receive data, where processing occurs and who has administrative access. Do not assume a private runtime makes every external connection private.
Log the agent identity, run identifier, target system, requested operation, result and associated human approval. Avoid recording unnecessary sensitive content. Alert on unexpected permission changes, unusual destinations and repeated denied requests.
Before launch, require evidence of permission testing, credential rotation, revocation and approval enforcement. Our AI governance controls guide provides a broader framework for testing these safeguards.
Start with one controlled workflow
Choose a bounded task, assign its identity and prove the controls before expanding access. AI agent identity security is easier to manage when each workflow has a clear owner, narrow permissions and a tested stop procedure.
Tech Engine provides managed AI agents that integrate with existing systems, operate in a private Australian-hosted environment and retain human approval. To review a proposed workflow and its access requirements, call 1300 088 324 or email sales@techengine.au.
Want this applied to your business?
Request an AI Blueprint and we will map the processes worth automating first.
Request an AI Blueprint