Personally identifiable information (PII) detection
helps an organization identify sensitive values in traffic passing through its Oracle Integration 3 Model Context Protocol (MCP) Gateway. The policy can be scoped to the gateway, a particular MCP server, or a specific MCP tool. It can inspect requests, responses, or both, and it can block traffic or redact detected information.
The important design decision is not simply which detector to enable. You must decide what information needs protection, which fields are inspected, where enforcement happens, what should happen when a match is detected, and how to prove the control works.
Consider a fictional procurement assistant that checks purchase-order status through an OIC integration. Its request may contain a buyer's email address or a supplier tax identifier. Its response may contain supplier contact details. The business still needs the purchase-order status, but the AI client might not need to receive every contact detail. A carefully scoped PII policy can help separate those needs.
Analogy: Inspect the right part of the parcel
Imagine a mailroom that checks a particular type of information before a parcel leaves or arrives. The inspection rule must say which parcel, which contents, and which direction. A check that examines only the return address cannot guarantee that every sensitive item inside the parcel has been found. PII detection works similarly: the configured inspection boundary matters as much as the detector itself.
📑 In This Post
- Plan the PII protection boundary
- Select built-in and custom detectors
- Choose payload fields and identify blind spots
- Choose scope and traffic direction
- Choose blocking or redaction
- Work through a procurement configuration
- Test the policy and collect evidence
- Plan an enterprise rollout
- Avoid common configuration mistakes
- Frequently asked questions
- References and further reading
- Summary
01
Plan the PII Protection Boundary Before Configuring Detectors
PII detection is a security policy, not a substitute for understanding the data contract of an MCP tool. A detector can identify certain patterns or information types, but the organization must decide which data is sensitive, where that data travels, and what should happen if it is detected.
For an MCP request, the direction is from the MCP client through the gateway toward the selected MCP server. For a response, the direction is back from the MCP server through the gateway to the client. The server might expose an OIC project integration or a separately managed third-party MCP tool. The PII policy operates on the traffic covered by its configured scope and enforcement point. It acts when traffic crosses that gateway boundary. If a user or client has already sent PII to an upstream model, chat history, or external host before the tool request reaches OIC, gateway request scanning cannot undo that earlier disclosure. Likewise, response scanning protects the gateway-to-client response path, not information already stored by the server or another system.
Console orientation — find the policy configuration area
Start here: The Oracle Integration instance and the project that contains the MCP Gateway configuration. Use a nonproduction project for any later changes or tests.
Navigation: Projects > select the relevant project > left toolbar MCP Gateway > Security policies section.
Read-only orientation
- Open Projects in the Oracle Integration navigation pane and select the project that owns the intended gateway. Do not assume the project that exposes an MCP server must be the same project in which your organization manages gateway policies; establish the actual object ownership first.
- Select MCP Gateway from the project's left toolbar. Identify the Security policies section and inspect existing policy names and descriptions, where your permissions allow.
- Record the actual gateway, server, and tool names shown in the project. Do not create or save a new policy during this orientation exercise.
Expected result: You have identified the project and gateway configuration area without changing an existing policy. Oracle's published policy-creation procedure is linked in Create Security Policies for MCP Servers.
The PII policy editor also requires a compartment OCID associated with the availability of the OCI Language service. OCI Language's own IAM documentation lists the permission AI_SERVICE_LANGUAGE_PII_ENTITIES_USE for PII detection API use. The public OIC policy procedure does not fully explain which runtime principal is used for every gateway-to-service interaction. Have the OCI administrator confirm that the identity and policy scope used by your configuration can invoke the required service; do not solve an access error by granting broad tenancy permissions. The Language permission name is not itself a complete IAM policy and does not establish that granting it to the console user authorizes the gateway runtime. Confirm the real principal and compartment. Treat OCI Language as a data-processing dependency: review what text the actual configuration submits, where it is processed, and which operations or traces may retain it.
Before selecting a detector, make a small data inventory:
| Question | Procurement example | Why it matters |
|---|---|---|
| What is sensitive? | Email address, telephone number, tax identifier | Identifies detector candidates and custom rules. |
| Where does it appear? | Tool input and tool result | Determines request-side and response-side protection. |
| Does the workflow need it? | Purchase-order status needs the PO number; it may not need a contact's phone number. | Helps choose between blocking and redaction without breaking the business task unnecessarily. |
| What is outside the inspection boundary? | Unselected fields, another tool, direct backend access | Makes residual exposure explicit. |
Start from the data contract and the business requirement, not from the list of detectors. A configured detector is useful only when the information it detects and the traffic it examines align with the risk you need to reduce.
A team knows that sensitive information might pass through MCP tools but has not yet documented its fields, traffic directions, or intended handling rule.
02
Select Built-In PII Detectors and Design Custom Patterns
A detector defines what the policy attempts to recognize. Oracle Integration's documented PII policy flow provides a Detection tab. In that tab, the configuration includes a compartment OCID for the OCI Language service and a selection of built-in detectors. Oracle's examples include email and telephone or fax numbers (US), social security numbers, and other detector choices.
OCI Language's PII documentation describes additional entity types, including names, addresses, dates or times, email addresses, some country-specific identifiers, bank-account information, routing information, card numbers, and certain credentials or secrets. Do not assume every OCI Language entity type is necessarily displayed or configured identically in the OIC Gateway editor. Use the choices visible in the applicable instance and verify their behavior. A service detector for bearer-token or JWT-shaped text does not establish that OIC inspects every Authorization header, transport credential, or protocol metadata field. Keep authentication material out of tool payloads and do not alter protocol/authentication fields as a redaction workaround. The OCI Language documentation for its pretrained PII identification and de-identification model lists English as the supported language; test your intended language and data before relying on results for multilingual traffic.
| Detector approach | What it helps recognize | What to verify |
|---|---|---|
| Built-in detector | A supported category such as email or telephone information. | Whether the relevant detector is selectable, recognizes representative values, and behaves acceptably on near-matches. |
| Custom pattern | An organization-specific identifier or a format not adequately covered by the built-in choices. | Pattern syntax, contextual matching support, false positives, false negatives, and the effect of the match on the chosen action. |
When is a custom detector useful? Imagine a company needs to recognize a tax-identifier format used by its suppliers. The built-in detector choices available to the gateway do not offer a suitable category. A custom pattern may help identify the expected character structure. That pattern must be tested against examples from the organization's data contract before the policy is used to block or transform real traffic.
Tricky concept: Matching a format is not validating an identity
A pattern can recognize a string that looks like a tax identifier. It does not prove that the identifier was issued, belongs to the person or organization named, or is valid in the relevant jurisdiction. Detection and authoritative business validation are different controls.
Illustrative pattern candidate—not a verified OIC editor entry: for an Indian Permanent Account Number (PAN)-shaped value, a proposed regular-expression candidate is shown below. The sample value is synthetic. This is an example for reviewing the idea with a security engineer, not proof that the value is a valid PAN or that the exact syntax can be entered unchanged into every OIC Gateway release.
Illustrative pattern candidate
[A-Z]{5}[0-9]{4}[A-Z]
Synthetic candidate: ABCDE1234F. This pattern tests a shape; it does not establish the identifier's authenticity or status. As a search pattern it can match a substring inside a longer code. Test embedded matches, adjacent letters/digits, lowercase text, whitespace, repeated values, and Unicode near-matches. Decide whether boundaries are required, then verify the syntax supported by the actual editor. Do not add whole-text anchors blindly when the goal is to find an identifier inside a sentence.
The gateway documentation states that an unavailable detector can be added through Add Pattern. However, the public OIC procedure does not enumerate every field, allowed expression syntax, contextual matching option, or save behavior inside the custom-pattern editor. OCI Language documents a separate custom PII-rule model with options such as a regular expression and contextual terms, but that does not establish that every underlying rule attribute is exposed or accepted by the OIC Gateway editor.
Therefore, inspect the actual Add Pattern editor in your release before finalizing an expression. Use its displayed input fields and validation behavior. Do not paste a complete OCI Language API rule document into a gateway field unless that form explicitly supports it. If the UI does not reveal the needed options, record the gap and confirm supported behavior with current Oracle documentation or Oracle Support.
Safety Check — custom detector quality
Risk → An overly broad pattern can flag harmless business codes, while an overly narrow pattern can miss sensitive identifiers.
Control → Test positive samples, negative samples, near-matches, and contextual examples. Make the intended identifier category and test results part of the policy record.
Remaining risk → A pattern match is not proof of authenticity, and test samples cannot represent every future format or maliciously crafted value.
The sensitive-data category is known and you need to decide whether an available detector is sufficient or a custom pattern requires controlled testing.
03
Choose Payload Fields and Make Unscanned Fields Visible
Field selection determines which parts of a payload are included in inspection. It is a separate decision from selecting a detector. For example, enabling email detection does not establish that every property in every tool request and response is inspected.
Oracle's policy-creation guide documents an Inclusion List when the tool-targeted option is selected. It states that only selected payload fields are inspected. This is a critical distinction: field-level inclusion must be documented as part of the policy's intended coverage, not assumed from the existence of a detector.
Begin by obtaining the actual MCP tool contract. The selected tool's identifier and input schema tell you which request properties can be sent. Its actual result schema or observed nonproduction response shows what may return. Names in the following table are invented for this example and must be replaced with the fields offered by the real tool.
| Traffic side | Illustrative field | Inspection decision |
|---|---|---|
| Request to the tool | requesterEmail |
Include if it exists in the actual input contract and email detection is required. |
| Request to the tool | supplierTaxId |
Include if present; pair the field with the appropriate built-in or custom detector. |
| Response to the client | supplierContactEmail |
Include in response inspection if the client should not receive the unredacted contact address. |
| Response to the client | supplierContactPhone |
Include if the phone number is sensitive and the tool can return it. |
| Request or response | freeText |
If it is present but excluded from inspection, record it explicitly as an unscanned field and test the consequence with synthetic data. |
Build an inclusion and exclusion record. In addition to listing the fields to scan, record relevant fields that remain outside the selected inspection set. An omitted field is not protected merely because it belongs to the same JSON object as a scanned property.
Worked example: An overlooked free-text field
Suppose the synthetic tool input includes purchaseOrderNumber, requesterEmail, and freeText. The policy includes the email field, but not the free-text field. A synthetic email address placed inside freeText illustrates a coverage gap: it may not trigger the intended field-scoped control because the field was not selected.
The correct response is to review the data contract and decide whether that field should be included, removed from the tool contract, or handled through another supported policy. The demonstration is a proposed nonproduction test, not an observed gateway result.
Be especially careful with nested objects, arrays, alternate response shapes, and fields introduced by later tool-schema changes. The public policy procedure does not fully specify how every nested or dynamically shaped property is represented in the inclusion selector across releases. If a sensitive value can occur in a nested object or list, inspect the selector and test that exact shape; do not infer coverage from a parent-field name. Also test duplicated values: the same email may appear in a contact field, a free-text message, an error, or another representation of the tool result. Inspect the complete client-visible result, not just one JSON property. Do not assume coverage for encoded content, attachments, images, or dynamically added fields when that behavior has not been established for this policy.
Safety Check — unscanned data
Risk → Sensitive information placed in an omitted field may pass without the expected inspection.
Control → Tie the inclusion list to the actual input and output contracts. Run a harmless synthetic-value test against each selected field and at least one deliberately omitted field.
Remaining risk → A field inventory becomes stale if the tool schema changes or a downstream system returns a new response shape. Revalidate coverage after relevant changes.
You need to prove which tool properties are inspected rather than relying on a broad statement such as “PII detection is enabled.”
04
Choose Gateway-Level, Server-Level, or Tool-Level Scope
The scope answers which portion of MCP traffic the policy targets. The enforcement point answers in which direction the policy is evaluated. They are different settings and must be documented separately.
| Scope | Meaning | Typical reason to choose it | Design consideration |
|---|---|---|---|
| MCP gateway | Targets traffic that passes through the configured gateway. | A common requirement applies to traffic across the gateway. | The policy is broad. Confirm whether every affected server and tool should share the same rule. |
| MCP server | Targets traffic associated with a selected MCP server. | One server has a specific data-protection requirement. | Other servers in the gateway may need separate coverage. |
| MCP tool | Targets a particular tool on the selected server. | A tool has a known sensitive input or response contract. | Tool selection and the applicable Inclusion List must match the real tool schema. |
For a first procurement design, tool-level scope is useful when the requirement names particular fields in one specific tool. A gateway-wide rule is more suitable when the same handling requirement genuinely applies across the gateway. A server-wide scope can make sense when all tools managed by that server follow a consistent policy.
Do not select the broadest scope just because it sounds safest. A policy that blocks required business traffic can cause a service outage; a policy that is too narrow can leave other tools exposed. The correct choice follows the data map and the intended enforcement boundary.
Request only means inspect the request before the gateway sends it to the MCP server. Response only means inspect the reply before it is returned to the MCP client. Both applies the inspection point in each direction.
| Enforcement point | What is inspected | Question to answer |
|---|---|---|
| Request only | Data moving from the client toward the MCP server. | Should the backend receive the sensitive field at all? |
| Response only | Data moving from the MCP server back toward the client. | Does the client need the original value to complete its task? |
| Both | The request path and the response path. | Are the privacy requirement and handling rule appropriate in both directions? |
The direction has practical consequences. A request-side rule can stop a sensitive value from being sent to the MCP server, depending on the configured action and detected match. A response-side rule acts on the returned traffic before it reaches the client; it does not undo the backend's earlier processing of that data. Selecting both directions does not imply that every log, credential, internal server field, or other route into the target system is also inspected. In particular, a response Block can withhold a result after the backend has already executed. It is not a transaction rollback. For any later write-capable tool, a blocked or failed response must not be interpreted as proof that no business change occurred or that a retry is safe.
Analogy: Two checkpoints, two different jobs
Checking outgoing parcels helps prevent selected contents from being sent. Checking incoming parcels helps control what recipients receive. A checkpoint on only one route cannot be treated as proof that the other route has been checked.
Field-selection caveat: Oracle's current procedure specifically documents the Inclusion List when a tool-targeted scope is selected. Do not assume that the same field-selection controls, fields, or selection semantics are available identically for gateway-level and server-level policies. Check the actual options exposed by the chosen scope in your release.
You must decide whether to apply a common rule broadly or protect a specific tool and its defined input or output fields.
05
Choose Between Blocking and Redacting Detected PII
Detection identifies a potential match. The primary action determines how the gateway policy should handle traffic when PII is detected. Oracle documents two primary choices: Block and Redact.
| Action | Documented behavior | Best question to ask |
|---|---|---|
| Block | Prevents traffic containing detected PII at the configured enforcement point. | Should this request or response be prevented from continuing if a match is found? |
| Redact: Mask | Masks detected values with a selected character, with options controlling which side is masked and how many characters remain unmasked. | Must any portion of the original value remain visible for a legitimate purpose? |
| Redact: Replace | Substitutes a configured replacement value for detected PII. | Can the consumer work with a clearly marked replacement instead of the original value? |
| Redact: Remove | Removes detected PII; the documented procedure requires no additional fields for this mode. | Can the workflow remain useful after detected content is removed, with the resulting value and payload structure verified? |
For Mask, the policy editor documents a masking character, the FROM_END or FROM_START unmask option, and the number of characters to leave unmasked. These are distinct choices. A masking character by itself is not a complete masking policy; the direction and permitted visibility affect what remains readable. Verify actual output for both direction settings instead of inferring which characters remain from the label alone. Test short values and cases where the permitted unmasked count equals or exceeds the match length. Partial masking can leave identifying information visible; it does not establish irreversible anonymization.
For Replace, decide on a replacement value that clearly indicates the original content is no longer present. An illustrative value is [PII_REDACTED]. The exact value is a design choice for the configured field and consumer, not a required Oracle constant.
Use Remove only after confirming that removing the detected substring will not make the returned data confusing or leave an invalid value where the business contract requires one. Redaction can protect information but may affect an integration's schema expectations or the client's interpretation of a tool result. Oracle's procedure does not define Remove as deleting an entire JSON property. Do not assume the key disappears, becomes null, or remains a schema-valid value. Observe the transformed content and validate the result against the real consumer contract, including required fields, formats, and enumerations.
Safety Check — a policy that changes business data
Risk → Redaction can remove information that a backend or client requires, while an inappropriate mask can still expose characters that policy owners expected to conceal.
Control → Choose the action separately for each relevant traffic path. Test original and transformed values, confirm schema validity, and verify that the useful business result survives.
Remaining risk → Detection may miss sensitive data or flag non-sensitive data. Redaction also does not establish that other services, logs, or direct routes have been protected.
The team needs to decide whether a match should stop traffic or transform the matched value while allowing the business interaction to continue.
06
Worked Configuration: Protect a Procurement Status Tool
This example turns the decisions above into a documented policy design. It is a worked design exercise, not a live lab. The tool identifier, property names, selected compartment, detector options, and test outcomes below are fictional or proposed until verified in an actual nonproduction instance.
Worked example: Procurement support
Business need: let an assistant retrieve the status of a purchase order without unnecessarily disclosing supplier contact information. Start with a read-only status tool. Do not include purchase approval or write operations in this example.
Assume the real tool's discovered schema contains fields corresponding to a purchase-order number, an optional requester email, and an optional supplier tax identifier. Assume its response may contain a status, a supplier email, and a supplier telephone number. Replace these sample field names with the actual tool contract before configuring any policy.
The design has two different intentions: stop an unnecessary sensitive identifier from reaching the tool in a request, and replace contact details in a response so the assistant can still receive the purchase-order status.
First, make the boundary explicit. The following table documents the intended configurations. Names are illustrative labels for the organization's policy register; they are not required Oracle identifiers.
| Design item | Request policy | Response policy |
|---|---|---|
| Illustrative policy name | PII-PO-REQUEST |
PII-PO-RESPONSE |
| Scope | MCP tool: the actual purchase-order status tool on the selected server. | MCP tool: the same actual tool on the selected server. |
| Enforcement point | Request only. | Response only. |
| Fields intended for inspection | The real input-schema fields carrying the optional requester email and supplier tax identifier, if those fields exist. | The real result fields carrying supplier email and supplier telephone information, if those fields exist. |
| Detector plan | Email detector and a tested custom pattern for the agreed tax-identifier format. | Available email and telephone detectors appropriate to the data. |
| Primary action | Block a request when configured PII is detected, because the intended status lookup does not need that sensitive content. | Redact and replace detected contact values using a clearly marked replacement. |
| Expected business outcome | A clean status request can proceed; a request containing a detected prohibited value is expected to be stopped at the configured enforcement point. | The client can receive the status with configured contact values transformed, if the response rule detects and handles those values as expected. |
Important configuration boundary: use the actual fields and options displayed in your tool-targeted policy. The public OIC procedure does not guarantee that a particular example property, nested path, or custom-rule option will appear in your tenant. If the schema exposes different properties, revise the policy record rather than inventing a field or treating a nonmatching name as protected.
Illustrative tool input payload. This is a business payload fragment, not a complete MCP wire exchange. It is not intended to be copied as a runnable protocol request.
{
"purchaseOrderNumber": "PO-DEMO-1042"
}
The purchase-order number is a synthetic business key. It is not presented as PII. It may still be confidential business data and may require other controls.
For a clean input, the proposed result is that the request can reach the read-only status tool, subject to normal authentication, authorization, tool eligibility, and backend rules. That outcome is an expectation to test, not an observed result.
Proposed negative request test. Use only fictional values in a nonproduction environment. Put a detector-matching value in a field selected for scanning, such as a synthetic email address in the sample requester-email field or a test value matching the candidate custom pattern in the sample tax-identifier field. Keep the caller, endpoint, tool, and other required arguments valid. This isolates the PII rule rather than accidentally testing a token or schema failure.
For the request policy, expected behavior is that the configured Block action prevents the traffic from passing the policy's enforcement point after PII is detected. Do not assume a particular HTTP status or error string; record the actual gateway result, and examine available sanitized evidence from the integration or target to determine whether the tool executed.
Proposed response test. Seed a synthetic nonproduction supplier record whose response contains an email address and telephone number in the selected result fields. Invoke the status tool with a clean request. The expected outcome is that the status remains available while the configured contact values are replaced before the response reaches the client. Confirm the actual output; do not assume the policy preserves every field or output structure without testing.
Worked interpretation: What proves the configuration is doing the intended job?
For the request test, a gateway rejection alone does not prove PII detection caused it. Confirm that the caller was otherwise authorized, the actual selected tool exists, the arguments were valid, and the rejection is associated with the intended policy. Where available, check downstream execution evidence to verify that the blocked request did not trigger the tool's business work.
For the response test, verify the exact client-visible result and ensure the original matched contact values are not returned in that response. A successful response test does not prove that logs, alternate tools, or direct backend access are covered.
Console walkthrough: create the two planned policies
Start here: the gateway-management project, with a tested and saved managed MCP server, its real tool metadata, a confirmed OCI Language compartment and runtime access, and synthetic baseline records. Capture the existing policy assignments and clean tool result before making a nonproduction change. The two policy names below are example labels.
- Open Projects, select the management project, and select MCP Gateway in its left toolbar. Under Security policies, select Add, then PII Detection.
- For the request policy, enter a descriptive Name such as PII-PO-REQUEST, review the generated Identifier, and add a description. Select Add to open the policy editor.
- On Detection, select the verified compartment OCID where OCI Language is available. Select the required built-in detectors. Use Add Pattern for the custom tax-identifier detector only after confirming its supported fields and testing the expression; do not paste an OCI API rule document into an unsupported form.
- On Enforcement, select MCP tool, then the actual registered server and status tool. Select request enforcement. Use the displayed Inclusion List to select the actual input fields. Resolve any missing or ambiguous field representation before claiming coverage.
- On Action, set Primary Action to Block. Review Detection, Enforcement, and Action together, then select Save.
- Create a second PII Detection policy for PII-PO-RESPONSE using the same verified tool. Select the appropriate contact detectors, response enforcement, and the actual response fields offered by the editor. Choose Redact, then Replace, and enter the intended replacement value such as [PII_REDACTED]. Select Save.
- Reopen each saved policy and verify its server/tool association, traffic direction, selected fields, detectors, and action. If the editor cannot select the required result field or combine the planned policies, record the configuration gap and revise the design rather than treating an example property name as protected.
Expected result: two saved policy definitions corresponding to the request and response intentions. This is configuration evidence, not proof of effective runtime protection. The public guide documents the main controls; it does not establish every nested-field, custom-pattern, or policy-composition behavior for every release.
Attach the policies and establish the effective gateway path
- In the same project, open the intended gateway from the MCP gateways section. Under MCP Servers, confirm the registered server that owns the selected tool.
- Under Policies, select the applicable saved PII policies for that server. Preserve existing tool filters and business policies. Review Policy evaluation order and save the gateway configuration.
- Confirm the gateway's effective runtime state and follow any lifecycle steps required by the deployed release. The guide establishes initial activation; do not infer that every saved policy edit hot-updates an active endpoint or always requires the same reactivation sequence.
- For runtime tests, use Actions > Run on the intended gateway to obtain its actual gateway URL. Connect the already authorized MCP client to that URL, not the source project MCP URL. Run the clean baseline, request Block test, and response Replace test separately.
If another attached policy rejects a request first, that result does not prove the PII policy ran. If one policy transforms a value before another evaluates it, the later policy may see transformed content. Review the chosen evaluation order, isolate the PII behavior with synthetic cases, and record which control actually handled the request or response.
Saving a security-policy object is not by itself proof that the intended gateway path is enforcing it. Verify the actual gateway configuration and relevant policy assignment using the current gateway documentation. If a policy or gateway change requires the gateway to be reactivated, follow the documented lifecycle and re-test after the change; do not treat an earlier successful test as proof that a later configuration is active.
You need a concrete design record that maps each sensitive field to its detector, traffic direction, scope, action, and expected evidence before touching a nonproduction gateway.
07
Test Detection, False Positives, and Unscanned Fields
A configured policy is a hypothesis about how a gateway will handle traffic. The tests must establish both what it catches and where its protection stops. The matrix below is a proposed test plan; none of these cases has been executed in your OIC tenancy.
| Test | Setup | Expected evidence |
|---|---|---|
| Clean request | Authorized caller, valid tool and arguments, and no detector match in selected fields. | Expected successful read-only business result, subject to other applicable rules. |
| Built-in detector match | Synthetic email or telephone value in a field selected for inspection. | Evidence that the expected policy action occurred at the intended request or response boundary. |
| Custom pattern match | Synthetic value matching the approved test pattern in the selected field. | Evidence that the pattern and its handling action work together as intended. |
| Near-match / false positive | A synthetic string that resembles an identifier but is not meant to be sensitive. | Record whether it is incorrectly detected and how that affects legitimate processing. |
| Unselected field | A harmless synthetic email placed in a field intentionally omitted from the Inclusion List. | Evidence of the policy's coverage boundary. A possible pass-through exposes a design gap; it is not a security success. |
| Response redaction | A test record returns sensitive values in selected response fields. | Client-visible output no longer contains the original matched values, while required status information remains usable. |
Add coverage and service-failure cases
Repeat detector tests in nested objects and arrays, in free text, with multiple matches, and on each supported response shape. Include representative language and country formats: a US telephone detector is not evidence of Indian or other international-number coverage. Verify mask direction and visible-character counts, Replace output, and Remove output as separate cases. Check every returned representation for the original synthetic value.
In a disposable nonproduction setup, have the responsible owner assess a detection-service authorization error, unavailability or timeout, and representative large payloads. Do not revoke shared permissions to manufacture a failure. Record whether traffic is blocked, errors, or proceeds, and decide whether that observed behavior satisfies the intended protection requirement. The cited OIC page does not establish a universal fail-closed, retry, truncation, or payload-size contract. Do not copy OCI Language API limits into an OIC gateway limit claim without verification.
Run negative tests carefully. If you are testing the PII control, do not use an invalid token, incorrect scope, nonexistent tool, missing required argument, or malformed unrelated value. Those faults can make a request fail without proving anything about PII enforcement. Ensure the test sends a fresh request through the changed gateway rather than displaying a cached response. Record available request identifiers and timestamps. Corroborate a request-side rejection with policy and downstream evidence; the absence of an integration-monitoring row alone is not conclusive.
For a request Block test, verify the rejection layer and, where evidence is available, that no unintended backend work occurred. For a response-redaction test, inspect the returned result with synthetic data and check for disclosure of the original matched value. A request not reaching the backend says nothing about whether a response-side rule works. Likewise, a successful response test does not prove the request-side rule works.
Keep evidence minimal and sanitized. Record the nonsecret policy identifier, test case, tool identifier, direction, expected and observed outcomes, sanitized gateway result, relevant execution evidence, and a status of passed, failed, or not executed. Do not paste production PII, bearer tokens, client secrets, signing material, or unmasked response bodies into screenshots or a test report. Gateway Production tracing is documented without request/response payload capture; Audit tracing captures payloads and is not recommended for production. Neither setting proves that client consoles, integration traces, OCI service logs, or target logs are redacted. Use synthetic data and review retention/access at each relevant layer.
Safety Check — evidence and interpretation
Risk → A request can fail for authentication, authorization, tool-routing, schema, service, or business reasons unrelated to PII detection.
Control → Keep all inputs valid except the one deliberately changed for the test. Identify the actual enforcement point and corroborate it with available sanitized evidence.
Remaining risk → The evidence demonstrates tested cases only. It does not establish perfect detection, coverage of untested formats, or protection for traffic that bypasses the gateway.
You need evidence that distinguishes a real PII-policy result from a generic tool error and makes known blind spots visible to reviewers.
08
Enterprise Rollout: Assign Owners and Preserve the Protection Boundary
PII protection for MCP traffic is a shared responsibility. The gateway administrator owns the policy configuration and its applicable enforcement boundary. The integration or MCP-server owner understands the business payload and the target workflow. The security and privacy teams define the handling requirement. The OCI administrator confirms the service and identity prerequisites. The application or data owner decides whether a redacted or blocked value is acceptable to the business process.
| Control area | Responsible role | Minimum evidence |
|---|---|---|
| Data classification and intended handling | Data owner with security/privacy input | A reviewed field inventory and reason for blocking or redacting each field. |
| Detector configuration | Gateway policy owner | Detector selections, custom-pattern rationale, and test outcomes. |
| Service and IAM prerequisites | OCI administrator | Confirmed region, compartment, service availability, and appropriately scoped permission. |
| Tool contract and backend behavior | Integration or MCP-server owner | Actual tool schema, representative synthetic results, and confirmation that redaction preserves the expected workflow. |
| Testing, change and incident response | Gateway operations and security | Sanitized results, named owners, a retest trigger, and an approved rollback approach. |
Schema and policy changes require review. A tool may change its input schema, result fields, or behavior. That change can make the existing field inventory incomplete even if the gateway still accepts the policy. Revisit the Inclusion List and the tests when the tool changes. Also review the configuration after relevant gateway or policy lifecycle events and confirm the effective state before relying on the control. Identify every gateway using a shared policy before editing it. Retain the prior configuration and a consumer-impact record, restore only an approved known configuration if needed, and rerun the clean and detection tests. Restoring a policy cannot undo backend processing or recover a value already disclosed.
Do not confuse PII detection with data-loss prevention as a whole. Detection is only one part of the design. It does not automatically prevent direct access to a backend, grant the caller the correct business permissions, establish user identity propagation, redact every operational log, or defend against prompt injection. Those risks need their own controls and evidence.
For consequential business data, keep authoritative authorization and transaction validation in the integration or target application where appropriate. The gateway's PII policy is an additional control at its configured traffic boundary; it is not proof that every other path into the target system is governed by that boundary.
PII policy changes will be maintained by several teams or form part of an enterprise MCP security review.
09
Common PII Configuration Mistakes
Mistake 1: Choosing a detector and assuming all traffic is covered. Detector selection and payload-field selection are different settings. Review the applicable scope and identify which input or response fields the policy actually scans.
Mistake 2: Treating the Inclusion List as a harmless performance choice. A field excluded from inspection can contain sensitive information. Record the intentional omissions and decide whether they represent accepted residual risk or a gap that must be closed.
Mistake 3: Applying response protection and assuming requests are protected too. Request and response enforcement are separate choices. Test the intended direction explicitly and use the combined option only where both paths need inspection and the behavior is acceptable.
Mistake 4: Using a broad pattern without testing false positives. A string that matches a format might be a legitimate business code. Test near-matches as well as valid synthetic examples before choosing a blocking action.
Mistake 5: Believing redaction validates a value. Masking, replacement, or removal changes how matched data is handled. It does not prove that an identifier is authentic or that it is permitted in the business process.
Mistake 6: Treating a successful save as proof of effective protection. Saving configuration does not replace a controlled test of the actual gateway path, field, direction, and action. Confirm the relevant policy configuration and execution evidence.
Mistake 7: Ignoring service and entitlement prerequisites. The PII editor's presence, the availability of MCP Gateway, and the OCI Language service's regional or IAM readiness are separate things to check. Verify the actual environment rather than inferring support from an example.
Mistake 8: Assuming the policy covers direct backend access. A client or integration that can reach a target through another route may not pass through the gateway policy. Review network paths, credentials, and target-side authorization separately.
Safety Check — before calling the design ready
Risk → A policy can look complete in the console while an untested field, unsupported pattern, or alternate route leaves a disclosure path open.
Control → Maintain a traceable mapping between data field, detector, policy scope, traffic direction, action, test case, evidence, and owner. Use actual tool schemas and synthetic tests.
Remaining risk → No detector or gateway policy guarantees perfect classification or closes paths outside its enforcement boundary. Production approval still requires the organization's security and operational review.
You are reviewing an existing PII policy and need to identify the most likely gaps before approving wider use.
10
❓ Frequently Asked Questions
Can an Oracle Integration MCP Gateway PII policy inspect both requests and responses?
Yes. Oracle documents request-only, response-only, and both-direction enforcement choices. Choose the direction according to the intended boundary and test the selected path separately.
Does enabling a built-in detector mean every payload field is scanned?
No. Detector selection and payload-field selection are separate concerns. Oracle documents an Inclusion List for the tool-targeted policy option and states that only selected fields are inspected.
When should a team create a custom detector?
Use a custom detector when an approved sensitive-data category is not adequately covered by the available built-in choices. Confirm the editor's supported pattern format, test false positives and false negatives, and do not treat a format match as proof of validity.
Should a team always block detected PII instead of redacting it?
Not always. Block is appropriate when traffic containing the detected information must not continue. Redaction may be preferable when the business operation can continue without the original value. Test the resulting payload and its effect on the business workflow before choosing.
Does PII detection guarantee that sensitive information cannot reach an AI client?
No. Detection can miss a value, a field may be outside the configured inspection set, or traffic may bypass the gateway. Test relevant data shapes and routes, use least privilege, review downstream access, and document residual risk.
11
🔗 References and Further Reading
The following primary Oracle sources were consulted for the product-dependent claims in this article. They establish the documented controls, prerequisites, and general behavior; they do not prove that a particular tenant's UI or configuration was tested.
- Create Security Policies for MCP Servers — Oracle Integration's documented procedure for the PII Detection policy, detector selection, scope, direction, field inclusion, and actions.
- Business and Security Policies for MCP Gateway — documented differences between gateway, server, and tool scope, along with the supported PII enforcement directions.
- Oracle Integration 3 Release Notes — September 2026 (26.10) — the published minimum release for the MCP Gateway feature.
- AI Feature Availability — regional availability information for the MCP Gateway and related AI features.
- Oracle Integration Editions — edition distinctions and entitlement considerations.
- OCI Language: Personal Identifiable Information — documented PII entity categories and custom PII-rule concepts.
- OCI Language Policies — OCI Language IAM policy concepts and permissions for PII detection operations.
- OCI Language Pretrained Models — supported model and language information for PII identification and de-identification.
- Create and Configure an MCP Gateway — assigning selected servers and policies, reviewing evaluation order, and saving the gateway.
- Activate an MCP Gateway — activation, inbound authorization choices, and tracing privacy.
- Get the URL for the MCP Gateway — retrieving the client-facing URL from Run details.
Technical documentation was reviewed on 18 October 2026. UI labels and product behavior may vary by release; confirm the release and actual editor options in your instance. The procurement data and expected outcomes are fictional teaching examples; no tenant result is asserted. Oracle and other vendor names are the property of their respective owners.
12
📝 Summary
Plan the boundary first. Identify the sensitive data and the request or response paths where it appears before choosing detectors.
Choose detectors deliberately. Use the supported built-in options where they fit; design and test custom patterns for approved data categories that need them.
Make field coverage explicit. Record what is included, what is excluded, and which unscanned fields create residual risk.
Choose scope and direction independently. Gateway, server, and tool scope answer where a policy applies. Request, response, and both answer which direction it inspects.
Choose an action that matches the business requirement. Blocking stops covered traffic when PII is detected; redaction transforms matched content through masking, replacement, or removal.
Test the control rather than assuming it works. Use synthetic values, valid baselines, deliberate negative tests, and sanitized evidence from the actual enforcement path.
For the procurement example, write down the actual gateway, MCP server, tool identifier, input and output fields, detector choices, scope, direction, and primary action. Then identify one valid test, one detector-triggering test, and one deliberately unscanned-field test. For each, state what evidence you would need before claiming the control passed.
Suggested checkpoint answer: use the real status-tool contract. Request-only Block targets the selected optional email/tax fields; response-only Replace targets the selected contact fields. A clean synthetic status request should remain usable. A matching request must be rejected by the intended PII policy before tool execution; a clean request returning synthetic contact details must reach the client with the intended replacements. A deliberately omitted free-text field demonstrates only the documented inspection boundary. Record all actual outcomes and leave these tests Not executed until run.
Keep the configuration tied to evidence, not assumptions. A useful PII policy is one whose data fields, detection logic, enforcement boundary, business impact, and tested limitations are all understood by the people responsible for it.

Comments
Post a Comment