Privacy and Security
Privacy of the Kilo IoT AI Assistant — session-scoped auth, permission inheritance, org isolation, data retention.
Authentication
Every interaction with the assistant is authenticated using your active session. The assistant inherits your exact permissions and organizational context. It cannot see more data than you can, and it cannot perform any action you couldn't perform yourself.
Permission inheritance
The assistant mirrors the access model of the account using it:
Organization administrators can query all organizational data through the assistant, matching the full visibility they already have in the platform interface.
Standard users see only the devices and data their permissions grant. If a user cannot access a device on the Devices page, the assistant cannot query that device either.
Read-only users can query freely, but the assistant cannot change anything on their behalf — it can only do what their role already allows. A user who cannot edit a rule in the interface cannot have the assistant edit it either.
Confirmation before consequential actions
When your role does allow changes, the assistant still asks before doing anything destructive or hard to reverse. Deleting a device or rule, or resolving an alarm, surfaces an explicit Confirm Action / Cancel prompt, and the assistant only proceeds once you approve. Routine, safe steps run directly; consequential ones are always gated by your confirmation.
Organization isolation
Conversations and data queries are strictly scoped to your current organization. The assistant cannot access data from any other organization, even if your account holds membership in multiple organizations. Switching organizations in the platform interface changes the assistant's data scope accordingly.
What is stored
Chat history: Your questions, the assistant's responses, timestamps, and session metadata are stored so you can revisit previous conversations.
Privacy: Chat history is private to your individual account. No other user in your organization can see your conversations with the assistant.
What is NOT stored
Raw device telemetry is not duplicated or retained by the assistant beyond the scope of your query.
Passwords, API credentials, and authentication tokens are never captured.
Billing and payment data is not recorded in conversation logs.
How queries are processed
Your question is sent to the assistant backend, authenticated with your active session.
A language model interprets the intent of your question.
The assistant queries only the data sources your permissions authorize.
The results of those queries are returned to the model, which composes a natural-language response from them.
The response is streamed back to your browser in real time.
Which data reaches the model
The assistant is agentic: it answers by calling tools against the platform and reading the results. That means the telemetry, device state, rules, and alarms it retrieves for your question become part of the model's context — that is how it can tell you which cold store drifted overnight rather than only describing how to find out.
Where that data goes depends on the model provider you choose:
The included allowance — requests are served through Kilo's configured model provider.
Your own model key (OpenAI, Anthropic, OpenRouter, or any OpenAI-compatible provider) — data goes to that provider under your own agreement with them.
A model you host yourself — set Base URL to your own endpoint and the data goes only there. Choosing the Ollama provider fills in Ollama Cloud, which is a hosted service like any other; self-hosting means replacing that address with your own, and the address has to be reachable from the platform.
Two limits hold in every case: the assistant reads only what your permissions already allow, and it stays inside your current organization. Raw telemetry is not duplicated or retained by the assistant beyond the scope of your query.
The same applies to the MCP server: when you connect your own AI client, the data it retrieves reaches whichever model that client runs. You choose the client and the model.
Best practices
Do not share passwords, API keys, or private credentials in the chat. The assistant does not need them and will not use them.
Be specific about devices and time ranges to receive precise, relevant answers.
All interactions with the assistant are logged for security audit purposes, consistent with the platform's audit trail.
Last updated