AI overview

Kutrion’s AI features help you operate the platform faster without handing your secrets or terminal output to a third-party model. The assistance is grounded in your own organization’s data and runs against a self-hosted model (local Ollama) by default. An organization may enable a cloud fallback, which needs a configured key, the organization’s opt-in and the cloud_ai entitlement; a prompt sent that way passes a secret-and-PII filter first. Which network the default model sits on depends on your deployment — see AI privacy model.

Command copilot

The command copilot turns natural language into a concrete command and tells you whether it is safe to run:

• Natural language to commands. Describe what you want and the copilot proposes the SSH or fleet command to do it — see Fleet commands.
• Safe / dangerous classification. Every proposed command is classified, deterministically, as safe or dangerous after the model returns it. A command flagged dangerous is gated behind a step-up before it can run, so the model’s suggestion alone never executes something destructive unchecked.

Command-history analysis

The copilot can analyze command history to support review — summarizing what was run and surfacing notable activity — so that going over a session or a stretch of work does not mean reading every line by hand. This analysis is grounded in your recorded activity, not in outside knowledge.

Access copilot

The access copilot answers questions about your environment, grounded in your organization’s own data:

• Grounded answers. Questions are answered from your audit trail, vault metadata and anomaly signals — for example, what is expiring, what looks risky, or who accessed what.
• Citations. Answers come with citations back to the records they are based on, so you can verify the claim rather than trust it. Retrieval is scoped to your organization, so the copilot cannot ground an answer in another tenant’s data — see Tenant isolation.

Rotation and access helpers

• Rotation helper. AI assistance helps with credential rotation, grounded in the same vault and audit data — see Sharing and rotation.
• AI-written JIT justifications. When you request just-in-time elevation, AI can help draft the justification that accompanies the request — see JIT access. The request still goes through the normal approval and audit path; the assistance is in the writing, not the granting.

Privacy by default

Every AI feature is subject to the same privacy model: assistance defaults to a self-hosted model rather than a cloud one, and the cloud path opens only when a cloud key is configured, the organization has opted in and its plan carries the cloud-AI entitlement. Secrets are stripped from anything that leaves, which is a filter on top of that decision rather than the thing making it.

The default says which kind of model answers, not whose network it runs on. On a self-hosted install the model is on your network because you point the server at it; on the hosted service the default model is the deployment’s own, and putting inference on your network is a configuration step with conditions attached. Read AI privacy model before enabling anything — it sets out both cases — and Self-hosted setup to run the model on your own network.

Where to go next

• AI privacy model — what the default guarantees, and what it does not.
• Self-hosted setup — configure a fully local install.
• Fleet commands — what the command copilot drives.