Security

Security

Built so that you can entrust work to AI with confidence.

Kanata is developed around the idea of a "harness": designing in advance what AI is entrusted with, which knowledge it may use, how its work is reviewed and recorded, and how it can be stopped. Here is how we handle data and permissions, and how AI usage is recorded.

Video: Security

An introduction to Kanata's security, including data handling, units for setting permissions, and how AI agent actions are reviewed and recorded.

The people shown are fictitious.

Data

Data handling

RAG governance

Managing the basis for AI answers

To use RAG on internal knowledge safely, it helps to decide in advance what AI may base its answers on.

Reference folders

Set the folders each AI agent refers to. Within those folders, only what the user in the conversation is allowed to view is used as the basis.

Answers within viewing permissions

Documents the user in the conversation cannot view are not used as sources.

Web search and instructions

Choose whether each agent uses web search. Things you do not want it to answer (prohibited topics) are set in custom instructions.

The setup steps are explained on the AI and RAG page.

Access

Units for setting permissions

Spaces and projects

Create projects for each purpose within your organization's unit (the space). You can build up shared assets while keeping each department's confidential information separate.

Roles per app

For each app, you can set member roles (such as admin and user) or mark members as "no access". In AppStudio apps, you set which screens and data each role can access.

Folder permissions

Permissions can be set for each folder in the Library. Members can view by default, and edit rights are granted per folder. The AI agent's reference scope is also limited to these permissions.

Harness

Controlling how the AI agent runs

Manage AI agents with a harness.

The harness is the mechanism that keeps Kanata AI Agent running safely.

  • Plan review: Writes and app or Studio executions run after the plan is reviewed according to the autonomy level. Publishing, overwriting, deleting and schema changes are confirmed even at autonomy level 3 (automatic execution).
  • Approval rights: Only the user who made the request can approve a plan. Only project admins can choose autonomy level 3.
  • Reference scope and execution permissions: Set the reference folders for each agent and the execution permissions for each tool (read, write, execute). Actions are always limited to the permissions of the requesting user.
  • Credit consumption: Credits are consumed as usage by the requesting user.
  • Credit management: Space admins check the remaining credits and purchase more as needed.
  • External posting: The agent does not post to external services such as Slack or Teams; if asked, it only provides a draft of the text.

Enterprise

Security measures on the ENTERPRISE plan

Security assessment at onboarding ENTERPRISE

We support security reviews when you adopt Kanata.

Documents

Terms and policies

FAQ

Frequently asked questions

QIs the data we enter used to train AI?

Kanata does not use your input data to train AI models. This is stated in our Privacy Policy.

QCan we separate what each department can see?

Create projects within a space, and separate members and permissions by project, app and folder. The AI agent's reference scope is also limited to these permissions.

QCould AI answer with content that is not in our internal documents?

The AI agent answers based on documents in its reference folders. How it should respond to questions not covered by the documents (for example, answering "I don't know") is set in custom instructions. Only agents with web search allowed refer to external pages.

QWhere are files stored?

Files such as recordings, training materials and documents are stored in the AWS Tokyo Region.

QWhat security measures does the ENTERPRISE plan offer?

We support a security assessment at onboarding.

QWhose credits does the AI agent use?

Credits are consumed as usage by the user who made the request. Space administrators can check the remaining credits and purchase more as needed.

Last updated: