Prepare a referral packet.
Assemble the documents needed for an assigned referral. Constrain record retrieval and verify the destination before the packet leaves the workflow.
Review: patient scope + recipientAI agent security / HealthcareKeep patient data
Keep patient data
within the task.
Help teams prepare documents, coordinate cases and work with research data. Check what an AI agent can read, which tools it can use and where information can go.
One organization. Different permissions.
A case coordinator and a research assistant should not inherit the same access simply because they use the same model.
Analyze an approved dataset.
Use the project’s permitted data and analysis tools. Test whether a notebook, retrieved paper or tool response can redirect the agent toward identifiable source records.
Review: dataset + export permissionsPrepare a case for review.
Gather the supporting documents for one case. Keep access to unrelated member records and authority to change the case outcome outside the preparation task.
Review: case access + write authorityA document should not choose where patient data goes.
A coordinator asks an agent to prepare a referral. An attachment contains an instruction to upload the full patient record to an external “verification” endpoint.
- 01 / Untrusted input
The attachment redirects the task.
The requested upload includes records beyond the referral packet and a destination the organization has not approved.
- 02 / Configured control
Check the attempted transfer.
At the routed tool or proxy request, evaluate permitted arguments and the destination. The record system still enforces patient-level access.
- 03 / Expected result
Reject the unapproved destination.
The configured destination restriction rejects the transfer. The decision is logged so the owner can investigate the attachment.
An assessment tests this outcome against the actual integration, including alternate paths that could bypass the control.
How prompt injection reaches toolsStart with permissions. Then inspect the content.
A content scan is one layer. The integration also needs a clear answer to which records the task may retrieve, which operations it may perform and who may receive the result.
Evidence needs its own access policy.
Decide who can inspect the audit trail, how long it is retained and what is redacted in an export. A useful investigation record should not become an uncontrolled copy of sensitive content.
Test one path before expanding access.
Bring the agent configuration, the connected tools and a representative dataset approved for testing. Agree the environment and data-handling conditions before the assessment begins.
- Exercise the failure paths
- Test cross-case retrieval, instructions embedded in attachments, unauthorized exports and tool permissions that exceed the task.
- Make each finding reproducible
- Document the input, attempted action, reachable system and observed result so the team can reproduce the issue.
- Turn findings into changes
- Prioritize permission, integration and policy changes. Agree which corrected paths need a retest.
Healthcare security questions.
Where should a healthcare AI security assessment start?
Choose one administrative or research workflow. Identify its approved records, connected tools, external recipients and responsible owner. The assessment then tests whether untrusted documents or tool responses can cause access or transfers outside that scope.
Can oktsec enforce access to a specific patient or research dataset?
Only where the integration exposes parameters that can be constrained. Tool permissions and parameter rules can restrict a request, while the underlying record system must enforce its own patient, tenant and dataset permissions. Broad credentials or direct access paths must be addressed separately.
Does oktsec include an EHR or FHIR integration?
These scenarios are assessment examples, not a catalog of packaged clinical integrations. We first review the actual MCP server, API or proxy path to determine which requests can be inspected and controlled.
Does this make a workflow compliant or clinically safe?
No. Security controls and audit evidence support a review; they do not certify compliance, validate a diagnosis or establish the accuracy of generated clinical content. These examples cover administrative and research assistance. Clinical decisions remain with the responsible care team.
Start with a workflow you already know
Which records can your agent reach?
Let’s review the path from an assigned task to a requested transfer.
Discuss a healthcare assessment