Cribl AI and Your Data
Cribl AI does not access all event data flowing through Cribl. Depending on the feature, it might access limited user-selected or sampled events, along with query, configuration, or operational metadata, to perform the requested or enabled workflow. Each feature accesses only the minimum context it needs when you (or an enabled automated feature) invoke it.
What Data Cribl AI Accesses
The following features access limited context when you invoke them:
- Cribl AI generates follow-up query suggestions by looking at your current KQL query, up to the last 50 query strings from your search history (but not their results), and sample events from the current search job, which Cribl might truncate for size limits.
- The KQL assistant helps create or refine queries by looking at the query currently in the search editor, metadata from up to 50 recently used Datasets (ID and description), and, when needed, sampled events from up to 3 Datasets with up to 5 events from each.
- The visualization assistant looks at the selected Dataset’s metadata, along with up to 5 sample events and inferred field information, to suggest effective ways to display the data.
- Pipeline Builder uses user-provided or captured sample events, subject to token limits, together with the Pipeline configuration and schema context loaded into the session, to help with transformation logic.
- The Billy chatbot can inspect your deployment configuration and live operational status when you ask, including Worker Groups, Sources, Destinations, Routes, Pipelines, Search and Lake metadata, health, and I/O metrics. Information that these read-only inspection tools return might be sent to your configured AI provider to generate a response. Some tools summarize responses or omit selected fields, and Global Variable values are excluded from chatbot inspection responses. Administrators can configure named sensitive fields to mask in Source and Destination API responses. Apply this configuration to each relevant Worker Group, and review it when configurations or integrations change. For details, see Ask About Your Cribl Deployment.
- Search investigations can run bounded searches, analyze results, and use Dataset Intelligence and Notebook context to help with an investigation. Optional web search is routed through the Cribl AI backend, which makes it unavailable in Cribl.Cloud Government.
- When you configure MCP integrations, Search investigations and the Billy chatbot can call tools on the external MCP servers you enable. Cribl sends the tool arguments the model generates to the external server, and the returned results can become context for the active AI model.
- Dataset Intelligence, which Search investigations uses, samples events and runs limited exploratory queries to build a Dataset reference card.
- Cribl Guard features might analyze user-selected or sampled events to recommend detection and masking rules. Background detection samples a subset of Pipeline events for on-Worker sensitive-entity detection using a downloaded model bundle.
- Other assistants, such as Notebook summary, git commit messages, and explain KQL or Pipeline, send the relevant query, Notebook section, or configuration diff for the specific action you invoke them on.
Session Persistence and Retention
Cribl saves Search investigation and Pipeline Builder sessions automatically as you work, including the conversation messages and the feature context for each session. Cribl retains a session for at least 30 days from its last activity, then deletes it. Sessions that you continue to use stay available past 30 days, because each update resets the retention period.
Model Training
The Share data to improve Cribl AI setting in AI Settings controls whether Cribl can use your Organization’s data to train Cribl AI. Whether the setting is on by default, and whether you can change it, depends on your plan:
| Plan | Setting | Can you change it? |
|---|---|---|
| Paid | Off by default. Toggle it on to opt in. | Yes. An Admin can turn it on or off at any time. |
| Free | On. Cribl uses data from your Organization to train Cribl AI models. | No. The setting stays on while your Organization is on the Free plan. To opt out, upgrade to a paid Standard or Enterprise plan. |
| Cribl.Cloud Government | Permanently off. The setting doesn’t appear in AI Settings. | No. |
Data collection for training does not begin before December 2026.
This setting is independent of other configurations. It doesn’t depend on your other AI Settings, whether anyone uses AI features, whether you use Cribl-managed models or a Custom AI Provider, or whether you have set up a provider at all. If data sharing is on, Cribl can collect in-scope data from your Organization even if nobody uses an AI feature.
When you share data, Cribl uses it to train the Cribl AI models that all customers share, which helps improve Cribl AI features. Every change to this setting is recorded in the audit log. See Audit AI Settings Changes.
What Cribl Collects for Model Training
When data sharing is on for your Organization, Cribl can collect your content, not only usage statistics. This includes:
- AI interactions: The prompts you submit, the task you’re trying to complete, and context such as sample events, configurations, and KQL queries. Also, the responses and artifacts that Cribl AI generates, and the edits or corrections you make to them.
- Agent and model traces: Prompts, responses, tool calls and their arguments, and tool results, along with model and product versions, timing, token usage, and execution status. Tool results can include your configurations and data.
- Stream event data: Samples of events that flow through Cribl Stream Pipelines, including event structure, field paths, neighboring fields, and Source metadata.
- Usage and outcome signals: How you interact with AI features and their results, such as accepting, rejecting, retrying, dismissing, or abandoning a result. Also, validation results, detection predictions, and the actions you take on detection findings.
- Derived data: Data that Cribl creates from the preceding categories, such as labels, verified answers, review scores, synthetic-value copies, embeddings, and retrieval-index entries. Examples that train decision models can also include decision questions, candidate options, scores, selected actions, and observed outcomes.
Event samples can contain personal or sensitive information, such as names, IP addresses, hostnames, health information, financial data, and government identifiers. Where feasible, Cribl will remove any personally identifiable information (PII) or sensitive data before using it to train Cribl AI. Cribl may need some raw PII to improve its models that detect sensitive data, including the models behind Cribl Guard. Before you turn on data sharing, check whether the data in your Pipelines is appropriate to share.
Cribl tags each record with the following information: a tenant ID, capture timestamp, source and lineage information, consent scope, permitted uses, and a retention rule. Cribl can also enrich in-scope records with limited account attributes, namely coarse segment and industry, and with support-case metadata. Cribl excludes account names, contacts, and commercial terms.
Storage, Retention, and Access
Cribl securely stores raw data, such as activity, AI and agent traces, and logs, and keeps it for 12 months after capture. Cribl stores processed training data in Amazon S3 and keeps it for 24 months after capture. Cribl automatically deletes data when the applicable retention period expires.
Only authorized Cribl engineers can access this data. Each access requires a documented request and approval, and Cribl grants it only for a defined period. Cribl manages, audits, and keeps a record of all access.
Opt Out and Delete Your Data
If you’re on a paid plan and turn off Share data to improve Cribl AI after a period of being opted in:
- Cribl stops collecting new data within 24 hours.
- Cribl prohibits the use of your data in any future training data.
- Cribl deletes the collected data across all stores within 30 days.
Deletion covers the collected data described in What Cribl Collects for Model Training. To request deletion of your Organization’s data, contact Cribl Support. Organizations on the Free plan can’t turn off data sharing, but they can still ask for their data to be deleted.
Cribl keeps minimal deletion records as evidence that deletion occurred.
How Model Training Relates to Other Data Programs
Telemetry: Cribl continues to collect, store, and handle the telemetry that it collects today. For more information, see Telemetry Data.
Custom Guard Model Training: The Share data to improve Cribl AI setting enables data sharing for the purpose of improving shared Cribl AI models, including the models that power Cribl Guard. It does not create a model specifically for your Organization. Custom Guard model training is a separate, assisted engagement with its own data and terms. To request it, contact your Cribl account representative and provide representative log samples and detection requirements. Turning on Share data to improve Cribl AI doesn’t request a custom Guard model. For details, see Request a Custom Detection Model.
Local Detection Versus LLM Processing
Background detection for Cribl Guard runs locally on the Worker Node using regex rules and a specialized named-entity-recognition (NER) model, not an LLM, so it does not send sampled events to an external chat-completion service. Only the separate Analyze detections workflow uses an LLM, to classify findings and suggest mitigations after detection. For the available detection models and how to select one, see Background Detection.
Cribl AI in Cribl.Cloud Government
Cribl.Cloud Government never contacts the Cribl AI backend, the Cribl-hosted service at ai.cribl.cloud that serves Cribl-managed models and related AI capabilities:
- No prompt, event, or configuration context leaves the deployment for Cribl-managed models. Every large language model (LLM) request goes to the Custom AI Provider that you configure and control.
- Cribl reports no AI usage analytics or telemetry from these deployments.
- Cribl never uses data from these deployments to train models. The model-training control is permanently off, and the Share data to improve Cribl AI setting doesn’t appear in AI Settings.
- Web search and Billy documentation search are unavailable, because both depend on the Cribl AI backend.
Cribl blocks these requests at the platform level, so a misconfiguration cannot route data outside the FedRAMP boundary. For the full set of differences, see AI in Cribl.Cloud Government.