Trusting an AI agent starts with understanding what its connected account allows it to do. “Help with my email” might mean reading selected messages, accessing an entire mailbox or sending messages as you. Those are materially different permissions.
Keep identity verification separate from a convincing message or voice. Our AI voice scam guide explains why a familiar-sounding request still needs an independent check before money or account information is involved.
Before connecting a service, translate the requested access into ordinary actions. Ask what the agent can read, change, share, delete or send, and whether those actions require another confirmation.
Read the connection screen
Use the application’s supported account-connection process. Check the organisation behind the application and each requested scope. Where possible, grant access to the specific mailbox, folder or project needed for the job.
Do not assume that a chat instruction narrows a broad technical permission. If the connected account can delete a file, telling the assistant to “be careful” is weaker than configuring the workflow so deletion requires a separately enforced approval.
Our personal AI agent guide explains how to start with a bounded task.
Separate reading from acting
For research and drafting, read access may be sufficient. Sending email, changing customer records or posting externally requires additional capability. Expand permissions when the task requires them and after reviewing how the previous stage behaved.
For a meeting workflow, begin with approved notes and draft output. Only add sending and task creation once recipients, owners and deadlines are reliably checked. See our email and meeting workflow for a practical sequence.
Define meaningful approval
A useful approval screen identifies the exact destination and action. For an email, show recipients, subject, body and attachments. For a purchase, show item, quantity, total and delivery details. For a file change, show the affected file and proposed modification.
Approval should apply to that action. If the plan changes, the user should see the change before the system proceeds. Business teams can assess this using our AI agent approval-bypass testing guide.
Treat external content as information
An agent may encounter instructions inside an email, webpage or document. That material can describe the task without becoming permission to change accounts or reveal information.
This is one reason connected agents need more than a well-written prompt. Our prompt injection methodology describes how businesses can evaluate whether outside content influences actions beyond the intended boundary.
Check logs and recovery
You should be able to determine which action occurred, which account performed it and what information changed. Check how to stop an active task and revoke a connection. Understand whether revoking access also ends already issued sessions or requires additional steps.
Keep recovery options for valuable records. Version history and backups can reduce the cost of an accidental change, but verify that they cover the relevant system.
A short access review
- List every service connected to the agent.
- Record whether each connection can read, write, share or delete.
- Identify the actions requiring a fresh confirmation.
- Check logs using a harmless test task.
- Disconnect unused integrations and review access when staff leave.
Does a familiar provider make every integration safe?
No single brand answers the whole question. The particular application, account settings, integration and workflow determine the access involved.
What should I do after an unexpected action?
Stop the task, preserve the relevant logs and revoke the affected connection where appropriate. Review what changed before resuming. Organisations operating connected agents can use our AI agent incident response playbook to plan containment and recovery. For an independent review of a connected deployment, explore Detox’s AI security testing and red teaming services.
Understand permission scopes
An integration may request scopes such as reading messages, sending mail, managing files or accessing contacts. Read access can still expose sensitive content; write access adds the ability to change or create records. Administrative access may affect other users.
Translate every scope into a scenario. “Read files” means the connected application could retrieve the contents of files covered by that scope. “Manage mail” may include deleting or modifying messages as well as sending them. If the connection screen is unclear, consult the provider’s documentation before accepting it.
Prefer narrow scopes and specific resources where supported. A project assistant should not require every shared drive if it uses one folder. A scheduling assistant may need availability rather than the full details of private events.
Connection methods matter
Use the official authorisation flow rather than sharing a password. Modern delegated access lets the account provider issue a token with defined scopes and gives you a place to revoke it later. Review the application name and publisher shown during connection.
Never paste one-time codes, recovery keys or full session tokens into a conversation. If an agent asks for them, stop and use the supported account interface. Store API credentials in the platform’s secret facility and restrict them to the intended environment.
For business systems, use a dedicated service account only when the design requires one. Record its owner, purpose and permissions. Shared credentials without individual accountability make investigation and removal harder.
Least privilege in practice
Start with a read-only, limited dataset and draft output. Add write access after the agent performs the task reliably. Add external sending only when recipients and content are shown for approval.
Time-limit access where possible. A temporary research project should not leave a permanent connection. Schedule periodic reviews and revoke tokens for abandoned tests.
Separate environments. An agent being developed against fictional records should not also hold production credentials. Use test tenants or sandboxes that reproduce the important permission model.
Design approval around the exact action
An approval should bind to the parameters displayed. If a user approves an email to one recipient, the system should not silently add another. If a purchase price or seller changes, seek approval again.
Avoid “approve all future actions” as a shortcut for high-impact operations. If a recurring task genuinely needs automatic execution, define allowed recipients, amounts, record types and time windows in technical policy.
Make rejection safe. The agent should stop or return to drafting without repeatedly asking or finding another route. Record approved, rejected and expired requests.
Email-specific controls
Limit mailbox access to required folders or labels where possible. Separate drafting from sending and require recipient review. Treat links, attachments and instructions inside messages as untrusted input.
Check how the agent handles forwarded chains. They may contain personal data or internal discussion unrelated to the current request. Define whether attachments can be uploaded to another service.
For automatic responses, create narrow templates and escalation categories. Complaints, account recovery, financial details and legal notices should follow the appropriate human process.
File and cloud-storage controls
Use a dedicated project folder rather than an entire drive. Test whether the agent follows shared links outside that folder. Define whether it can create public links, change permissions or delete versions.
Preserve version history for important documents. Before bulk renaming or reorganising, generate a preview containing old and new paths. Test rollback on a small set.
Document retrieval should respect each user’s access. An agent must not copy restricted content into a shared summary simply because its service account can read both locations.
Financial and commerce accounts
Researching a product, preparing a basket and authorising payment are different levels of access. Keep them separate. Define maximum amounts, approved merchants and whether subscriptions are permitted.
Require confirmation containing the final total, currency, merchant, item and destination. Check for delivery charges, tax and recurring terms. Use provider notifications and transaction controls where available.
Never allow an email or webpage to increase a spending limit. Policy changes should come through an authenticated administrative channel.
Logging without creating another leak
Logs should identify the user, connection, action, destination, result and approval reference. They should avoid storing passwords, tokens or entire sensitive documents when metadata is sufficient.
Protect logs against ordinary users modifying them. Define retention and access. Test whether administrators can reconstruct a harmless action from request through result.
Alert on unusual destinations, repeated failures, mass changes and activity outside expected hours. Tune alerts using normal operation so staff can respond meaningfully.
Offboarding and periodic review
When an employee leaves or changes role, remove their agent access and review connections they created. Transfer ownership of legitimate workflows to managed accounts. Revoke unused tokens rather than relying only on deleting a chat.
Quarterly or another risk-appropriate interval, compare the connection register with provider account pages. Remove orphaned applications and verify that each owner still needs the granted scopes.
Review again after a product adds features. A previously read-only assistant may gain action capabilities that change the risk of an existing integration.
Test with safe scenarios
Create fictional messages and files covering normal work, ambiguity and hostile instructions. Confirm the agent stays within selected folders, asks before external action and reports failures accurately.
Test a changed destination after approval, an expired approval and a retry after timeout. Check that revoking the connection prevents further access promptly.
For business deployments, combine functional testing with authorization, data-separation and prompt-injection assessments. Record exact evidence and retest after fixes.
Is read-only access harmless?
No. It can expose confidential content and enable summaries or exports. It reduces modification risk but still requires appropriate scope and data handling.
Should every action need confirmation?
Routine low-impact actions can run under well-defined policy. Consequential, unusual or externally visible actions should have proportionate controls. The decision should come from risk and business process rather than convenience alone.
Permission-review worksheet
Create one row for every connection. Record the account owner, integration, scopes, accessible resources, permitted actions, business purpose and review date. Add the revocation route and the person responsible for incidents.
For each write capability, define the allowed target and whether confirmation is required. Include maximum values for spending or bulk operations. State what happens after a timeout and how duplicate actions are prevented.
Run a harmless test and capture the user, requested action, approval and final record. Revoke the connection and verify a subsequent request fails. Repeat after role changes or major provider updates.
Keep this register with ordinary access-management records rather than inside one person’s chat. It should support offboarding, periodic review and investigation without depending on memory.
A practical approval conversation
Before connecting an account, the requester should explain the exact task, the minimum source data and the expected output. The approver should compare those needs with the scopes displayed by the provider. If the integration asks for every mailbox when one shared folder would suffice, stop and find a narrower method.
Record how the user will know an action succeeded and who handles an incorrect action. For external messages, file deletion, payments or account changes, require a preview that shows the final destination and material details. The approval must relate to that specific action, not a vague consent granted weeks earlier.
Finish by testing disconnection. Remove the integration, confirm its access fails and check whether stored copies or logs remain subject to a separate deletion process. This short exercise exposes dependencies before an incident or employee departure makes the same task urgent.