Skip to main content

FAQs and known issues

Documentation for Tutela v3.9.0.

Answers to common questions and known limitations, with clear next steps. A known issue describes an observed problem, not a claim that every deployment is affected.

Google Search unavailable​

Why can Gemini answer an ordinary question but fail when Google Search is enabled?

Status: Under investigation. Last reviewed: 2026-10-12. Applies to AI Chat using Gemini with Google Search, also called search grounding. Provider limits can change independently of your Tutela release.

In an isolated check using Gemini 3.8 Flash, Google accepted the model lookup and an ordinary chat request, but rejected the search-enabled request with 429 RESOURCE_EXHAUSTED. Ordinary answers and answers that retrieve current web sources use different provider capabilities; success with one does not establish availability of the other.

What we know: Google refused the search-enabled request. This observation does not establish an invalid API key, a Tutela policy block, or a failure affecting every Google project.

What remains unknown: The exact quota or account restriction has not been identified. Search allowances, request limits and billing restrictions need to be checked for the affected Google project. No successful recovery has been verified for this report.

What does this mean for my answer?​

A successful ordinary answer does not establish that information is current. When live search is unavailable, do not rely on an answer as verified current web information. Check the source status and the actual citations, not only whether text was returned. Do not disable inspection or policy controls to work around a provider search failure.

What should I do?​

If you use AI Chat: Record the time, time zone, selected provider/model and any visible request ID or error. Ask your administrator to check the provider account. Avoid repeatedly resending the same request while the cause is unresolved.

If you administer the provider account: You need permission to view the Google project associated with the configured Gemini credential. Review that project's Google AI Studio usage, Google Search grounding limits, billing status and spending caps. Consult Google rate limits and Google billing guidance.

Google applies rate limits per project, not per API key. Replacing a key in the same project does not reset its quota. Do not raise a spending cap or switch billing plans without your organization's approval. The appropriate recovery depends on the identified limit: a documented reset, an approved quota or billing change, or escalation to Google if the recorded usage does not explain the rejection.

How do we know it is working again?​

After the relevant restriction changes or a documented reset occurs, try one harmless, time-sensitive question with search enabled. Confirm that retrieval succeeds and returns usable source references, then inspect the cited pages. An ordinary answer, a successful connection check or a changed key alone does not establish that Google Search recovered. If search is still unavailable, preserve the new error and escalate rather than repeatedly retrying.

These are diagnostic recommendations, not a verified recovery procedure for this unresolved report. See Grounding with Google Search for the provider's feature description.

What should I include in a support report?​

Include your Tutela version, provider/model, problem time and time zone, visible error code and request ID if present, and whether an ordinary request works while search fails. An authorized administrator can supply quota or billing observations through your approved support channel. Do not share API keys, passwords, raw prompts, customer files or full responses.

Use Settings > Support & diagnostics if your role permits it; otherwise ask your administrator or contact Tutela support. Never paste a credential into a support ticket or chat.

What does a detection score mean?​

A detection score is one input to a policy decision, not a probability that an alert is correct or a guarantee that an action is safe. Different detection methods score differently. A high score can still be wrong, and a low score does not prove that content is harmless.

A policy's minimum score is a cutoff from 0 to 1. Older screens may show an equivalent percentage: 80% means 0.8, not an 80% chance of a correct result. A template's preset is not a universally safe recommendation.

The cutoff is only part of the match. The policy must also apply to the relevant information and scope; another stricter matching policy can determine the final outcome. A detection, a policy match and a confirmed execution are different facts.

If you see an unexpected warning or block, check the recorded detection, matched policy and version, and any missing evidence. Ask an authorized policy administrator to compare permitted and rejected synthetic examples before changing the cutoff. Raising it simply to remove a warning can also cause real detections to be missed. Keep the previous configuration and request or inspection IDs for support, without sharing sensitive content.

Why was an ordinary comparison question flagged?​

Status: Under investigation. Last reviewed: 2026-10-18. Ordinary comparison questions have triggered an injection warning in a reported AI Chat scenario. This does not mean every comparison question is affected, or that a warning proves an attack. A fix has not yet been verified for this report.

The Laya stable release includes prompt-injection and tool-injection detection. The October report above remains historical: installing the replacement does not prove that this particular false positive is fixed. Check the installed build and reproduce the reported example with synthetic content before claiming resolution.

For the earlier warning, check the recorded detection separately from the matched policy and version. The applicable policy determined the response. A warning, a blocked response and a failed request are different outcomes.

Record the Tutela version, provider/model, occurrence time and time zone, and request or inspection ID. Ask support to review a harmless synthetic example that reproduces the problem. Do not send customer prompts, files, full responses or credentials. Use Settings > Support & diagnostics if permitted; otherwise contact your administrator or Tutela support.

Do not raise a cutoff or disable a policy just to silence this warning. Such a change can also miss genuine threats and needs an authorized policy review. There is no verified general workaround for this report; repeated unchanged requests do not establish recovery.

Does a Finding mean an attack happened?​

No. A current Finding records a policy match selected for review or a supported protection failure. The queue is not an attack count. It does not independently prove an attack, a data leak or a successful business outcome. Ordinary Activity, pending approvals, tool changes and collection problems have separate workflows. Historical findings keep their original meaning; they are not reclassified as current policy matches.

Start with the affected application or agent, occurrence time and time zone, recorded result, and why the concern was raised. An external AI request does not automatically establish a data leak. A policy allowing it does not establish safety. Use a Case when investigation needs an owner and notes; a Case's status does not change the underlying execution record or policy.

Why can an assessment be incomplete when an action completed?​

Tutela may have an exact completion record while a wider source assessment could not retrieve or assess all relevant records. For example, an AI request may have completed and its response was returned, but a scan limit still prevents a full assessment of the separate concern. Keep the concern, recorded outcome and source health separate.

Check what the limitation actually identifies: a missing action record, missing attribution, or a source retrieval problem. The assessment time is not the action's occurrence time. Ask the source owner to examine the listed record references; ask an administrator if your role cannot open them. A connection change does not necessarily recover historical records. Include identifiers and times in a support handoff, not raw prompts, responses, tool arguments, files or secrets.

Validate​

  • Record the affected provider/model, request time and error without sharing sensitive content.
  • Identify the applicable project restriction with an authorized administrator. If it remains unknown, keep the cause under investigation.
  • Record recovery only after an actual search-enabled request succeeds with usable source references. Ordinary chat success is insufficient.

Next steps​

  • Check the affected provider project before retrying search.
  • Keep this issue open until recovery is verified; a documentation update does not resolve the provider limitation.