FAQ: security, masking and traffic analysis
40 direct answers about controls, decisions and common errors.
Use the contents or question search. Each answer links directly to the control and its detailed instructions.
How do I disable masking?
verdict_classes does not disable replacement; it changes rediscovery. Change the targeted policy action and verify the final decision, key mode and other active rules.
How do I disable traffic analysis?
Stopping refresh, feeds or alerts does not stop inspection. Choose the intended mechanism: policy action, Router key mode, export or content capture; there is no shared switch for them.
How do auto and verdict_classes differ?
Auto searches known private values again; verdict_classes uses initial verdict classes. Both can replace values. off/false compatibility flags mean verdict_classes.
What is KEY_N and why do placeholders change?
KEY_N is a replacement-dictionary token. Stability depends on masking session, owner, TTL and storage. Use the supported session_id for conversations; another or expired session can have a different dictionary.
Why did masking become a block or on-prem route?
The verdict may be unreplaceable: substituting values cannot remove confidential meaning. Containment escalates; an unavailable private destination can lead to blocking. Check reason and providers.
How do I handle a DLP or injection false positive?
Find request ID, verdict, point and winning rule; reproduce with a synthetic anonymized example. Refine the detector or narrow scope on staging and retest positive/negative cases. Do not disable all protection for one record.
Why is original text missing in audit?
Metadata can be stored without raw content. Check capture enablement, read permission, retention and the separate store. Enabling capture now cannot reconstruct previously uncaptured or pruned text.
Does zero-day retention disable audit logging?
No. Zero disables automatic pruning, not new decision logging. Masking-session TTL and original-content retention are separate.
Does Pause or Off turn protection off?
In reports these stop display/refresh. Pausing a Secure policy removes that policy from decisions while others remain. Check which object the control belongs to.
Why does a report show no events?
Check tenant, UTC window, filters, source and actual traffic. Ring buffers are size/restart-limited. A 503 means the source is unavailable, not that no threats exist.
What should I do on 401 or 403?
401: renew the session or check inference-key validity, expiry and revocation. 403: check role, tenant scope and edition entitlement. A /v1 key does not replace an admin console session.
Why do settings fail to save with 422?
The server rejected the document. Read the error for action, condition, scope or missing fallback. Compare JSON against the canonical contract; older form values can differ. Rejected writes are not applied.
Why is a section or edit button missing?
Check user role, edition/license and configured components. Documentation does not grant permissions or enable Preview runtime. Ask your organization administrator for the required access.
Does DLP continue if alerts or webhooks are disabled?
Yes. Alert rules count events and subscriptions deliver them. Inspection, policy enforcement and audit logging are separate mechanisms.
Why are webhook events not delivered?
Check enabled state, subscribed events, receiver URL and signature verification. Inspect HTTP response/errors in deliveries; ping/replay really send events. Alert generation does not prove receipt.
How do I configure classifier timeout or circuit breaker?
Operators configure these service parameters; Classifier Health reports state/history. Specify the desired value and unit, then qualify failure/recovery and request decisions on staging.
How do I enable or disable one policy?
Policies → select your document → Activate, or Edit document → Status: Active → Save policy. Pause temporarily excludes it; Activate resumes it. Hosted replicas reread shared PostgreSQL without manual reload. Check a fresh request for your tenant; other policies may still block.
How do I monitor matches without blocking?
Use flag in a narrowly scoped rule. It passes text with a marker; stricter active policies still win. Check the final action, not just the editor selection.
Why doesn’t allow override another policy’s block?
Across active policies the strictest action wins: allow < flag < mask < reroute < block. Within one policy first match wins. Inspect resulting policy_id/rule_id.
What happens when no rule matches?
With available inspection the result is allow; inspection failure is fail-closed. Cover intended conditions/scope and test clean, private and failure cases on staging.
How do I add a custom regex or keyword detector?
An operator mounts JSON patterns with id/name/check/severity and regex/keywords through IAM_SEC_CUSTOM_DLP_PATH. Custom DLP → Reload → Test. The reference includes a complete file and positive/negative examples. Then configure a rule for the class ID; failed reload keeps the previous set.
How do I inspect model responses and tool output?
Use point=response or tool_output in integration/tests and match rule point. Chat/completions proxy inspects its configured flow; standalone inspect returns decisions your app must enforce.
Does batch inspection change application traffic?
No. It is a corpus dry run without an LLM call. Review each result/error; applications enforce decisions separately. Batch does not enable interception of another app.
How do I enable or disable SIEM?
SIEM → Endpoint/Format → Enable SIEM export → Save. Test delivery is available when enabled. Disabling stops export while retaining audit/inspection.
Does resolving a false-positive incident remove a rule?
No. This changes investigation state. Change policy/detector separately and test a new request. The originally blocked request is not executed by changing incident status.
Why did an application-scoped policy not match?
The released version rejects nonempty scope.apps and scope.groups with 422 and disables their form fields because trusted enforcement bindings are unavailable. Use a specific scope.tenant. Older versions could save these lists without restricting enforcement; do not rely on them. Report filters and Context do not create isolation.
Should I enter 50 or 0.5 for the block-rate threshold?
Incidents → Correlator → Configure: enter 50 in Blocks %, but send 0.5 as API block_rate_threshold. The API accepts a fraction >0 and ≤1. Five blocks out of ten meet the threshold if the minimum sample is satisfied. Apply and read back.
Why did the correlator not create a new incident immediately?
Thresholds are evaluated on the next scan (source default: every 60 seconds). Check tenant, window, minimum sample and Dedup: the same rule+tenant is suppressed for one window. Up to 2000 recent entries are read; a green indicator does not prove a fresh scan. Correlation does not change request decisions.
Does the tenant filter isolate a custom DLP detector?
No. Patterns-file detectors are gateway-wide. Limit actions with a specific scope.tenant and test another organization. Nonempty scope.apps/groups are rejected with 422. A detector has check, severity and regex/keywords, but no sensitivity field.
Customer created but invitation missing — create again?
First refresh Customers administration: saving the account and sending email are separate stages. If the record exists, repair delivery and Send invitation from that record. Avoid duplicates. Notify about saved settings sends a separate message; enabling notifications alone sends nothing.
How do I temporarily deny and restore customer access?
Customers administration → customer row → disable Access enabled → Save settings. Reread and verify denial of the customer session/key. Re-enable, save and sign in to restore. Tenant filters and disabled email notifications do not deny access.
How do I confirm a setting actually took effect?
After Save, reread server JSON and send a fresh synthetic request: request ID, action and policy_id/rule_id. The hosted release of 15 October 2026 reads shared PostgreSQL policies on all replicas without manual reload. The operator checks every replica; an unavailable store does not acknowledge writes and inspection blocks. On alternating 200/404, cancel an empty form and provide ID and time.
Why is the retention field disabled?
Check supported/enabled in the serving gateway response. If the version requires a qualified archive and reports supported=false, pruning is unavailable even when old days are visible. The operator sets up the supported storage lifecycle. Original-content retention and masking TTL are separate; do not delete data for testing.
How do Alert Rules thresholds differ from correlation thresholds?
Alert Rules uses ≥1 event over ≥1 minute with cooldown ≥0, e.g. 10 DLP events in 5 minutes. Correlation uses bursts, minimum requests, block fraction and seconds, e.g. 50% of 10 requests in 600 seconds. One alerts; the other creates grouped incidents; neither configures request action.
How do I inspect multiple texts without missing an item error?
Batch inspection → Lines (one nonempty line per item) or a JSON id/text/point/context array, maximum 1000. Run inspection → expand each row → Export CSV. Check IDs, count and each item error; HTTP success does not mean all items passed. This is dry-run without LLM generation.
Where do I configure masking Redis, encryption and mapping TTL?
Masking shows Session backend, encrypted and TTL as read-only. The service operator configures them. Save here only changes auto/verdict_classes; it neither enables Redis nor migrates mappings. Verify actual storage and mask/demask on the required replicas.
Why does a row say block while the Blocked counter is zero?
The older console counted only the blocked alias, while the server returns canonical block. Check every row and its raw JSON rather than relying only on the summary card. In the console released to staging and production on 15 October 2026, the staging check showed Total 2 and Blocked 1: an ordinary question was allowed and an injection was blocked. This display fix does not change the server policy.
Why does the incident screen go blank after a refresh?
In the older console, an automatically correlated policy_violation incident could trigger a rendering error. The console released to staging and production on 15 October 2026 accepts that wire type and null evidence arrays. Opening and cancelling correlator settings was checked with an actual incident. If it persists, reload the page and give your operator the time, URL and incident ID without request contents; there is no need to change thresholds or disable protection.
Have replica persistence and masking switches been tested?
Yes. The server reached hosted staging and production after 642 gateway and315 shared-core tests with PostgreSQL plus live scenarios. Shared staging verified UI save and active → block → paused → active on both replicas without reload. A separate isolated stand retained policy/mode after replacing both pods and verified UI auto → verdict_classes → auto, an email becoming [CONTACT], and no receiver call for block. This does not establish Redis recovery after storage loss or every replacement class.
Does the Secure gateway need PostgreSQL superuser access?
No. The gateway uses explicit DML grants without schema ownership or administrative memberships. A separate administrator runs migrations, gateway startup uses IAM_DB_SCHEMA_MODE=verify, and a readonly identity creates backups. Customer API keys do not grant database access.