Samurai App (the "Service"), provided by Kiva Corporation ("Kiva," "we," or "us"), is an AI agent OS for enterprises.
In the Service, AI agents read your company's knowledge, remember the context of your work, compile figures and prepare documents, and connect to external business systems to carry real work forward. Staff simply make requests in chat, with no need to operate each screen one by one.
Because of this, the range of information the Service handles and the range of operations it can perform are themselves the central security questions. The Service therefore treats control over who can reach which data, and how far they can act on it, not as a feature but as a premise of its design.
SamuraiAX, also provided by Kiva, is a business transformation service that combines AX consulting with a business OS. The Service is the business OS within it. This document covers the Service only; it does not cover consulting or custom development services.
The Service is designed on the following four premises.
Conversation records, business data, accumulated procedures (skills), memories, and audit logs are held in an independent database for each customer company. Your data never sits alongside other companies' data in a shared database. The work procedures that grow as you use the Service also accumulate only within your own company.
Agents have no permissions of their own. Both data access and operations on external systems are always evaluated against the permissions of the person who made the request. Going through an agent never gives access to information that the person could not reach directly.
Operations with real effect, such as writing to external systems, require approval before they run. Operations left unapproved are not executed. Requiring human approval for committing actions such as finalizing amounts, making payments, and external disclosure is part of the standard design.
Business operations are recorded in an append-only audit log and kept in a form where tampering can be detected. You can explain who did what, when, and on what basis in internal audits and financial audits.
The Service is offered only as a cloud service; we do not install it in customer environments (on-premises or private cloud). On the other hand, there is no restriction on what it can connect to. It can reach core systems running on-premises as well as portals specified by your business partners.
The Samurai App Security Whitepaper ("this document") describes the information security measures and data handling for the Service.
The Service is intended for corporate customers who handle internal business data and information about their business partners. And because it is an AI agent, there are situations in which it performs real operations on customers' business systems.
So that such customers can use the Service with confidence, this document explains the security measures we take, our data handling policy, the controls applied when AI inference is used, and our operational structure.
This document is intended as a basis for security assessment and internal approval when you are considering adoption. We respond to your own security checklists separately from this document. Specific configurations, such as the connection method, how credentials are stored, and log retention periods, are finalized in individual discussion with your IT and audit departments.
The Service is a product under continuing development. This document describes what we do not cover in the same detail as what we do, so that you can judge it against your own requirements.
Each statement in this document reflects information as of August 31, 2026.
This document is provided solely for your information. You use this document at your own responsibility.
Information you obtain through this document is provided by us on an as-is basis and to the extent we are able to provide it, without any representation or warranty of any kind, whether express or implied.
With respect to your gathering of information about the Service, your consideration of using the Service, your use of the Service, and any other acts, this document does not form part of any contract between you and us. Where a contract between you and us exists, the relationship between you and us is governed by that contract, and that contract is not modified by anything stated in this document.
This document reflects our information as of the time of writing. The Service is under continuous development, and the contents of this document may change without prior notice.
Except where expressly agreed otherwise in writing between you and us, we accept no liability whatsoever for any damage arising from the use of information obtained from this document.
The Service controls the range of information AI agents handle and the range of operations they can perform, as follows.
First, the place where data is stored is itself divided into a separate partition for each customer company (a "tenant"). Conversation records, business data, accumulated procedures, memories, and audit logs are all held in an independent database for each customer company, and never sit alongside other companies' data in a shared database.
The environment in which agents run code is also a disposable space, separate for each conversation, and is discarded when execution ends. Even within the same company, users cannot see files that another user created partway through a task. File storage and search indexes sit on shared infrastructure but are divided into per-company partitions. The system assigns the partition automatically, based on the membership information settled at the moment the user's request is accepted. Any attempt to access data before membership is settled results in an error.
At the entry point, access can be narrowed in two stages: verification of the user and the network the connection comes from. An account does not become active until its email address is verified, and at sign-in we always require a one-time code sent by email in addition to the password. Users have no way to turn off the one-time code requirement themselves. Single sign-on supports OIDC and SAML (single logout is not supported), with one configuration per company. Because you prove ownership of the email domain with a DNS record, email addresses outside verified domains cannot sign in.
You can also set, per company, the network ranges from which connections are allowed. The check uses values that cannot be rewritten from the client side, and if the check cannot be made, the connection is refused. However, this restriction only prevents reaching the company's data. The sign-in page itself can be opened from outside the allowed ranges, because at sign-in the user's company has not yet been determined and the check cannot be made.
Agents have no permissions of their own; both data access and operations on external systems are evaluated against the permissions of the person who made the request. Access rights to knowledge and business data can be set per user and per group, and search results are filtered by those rights as well.
As an exception, there is a method in which we hold permissions for an external system on a per-task basis. In this case, the credentials are bound to the task rather than to a user. For example, an employee who has no access to the accounting system can still perform only the "register an invoice" operation that an administrator has allowed. The user can use only that narrowly defined operation and never receives the credentials themselves. Both the person who ran it and the credentials used are recorded in the audit log. Note that the operations the Service can perform on external business systems are limited to the permissions of the accounts you issue to us.
For AI inference, we do not own a large language model of our own; we use models from external AI vendors. All calls go through routes we manage, and nothing is sent directly from your browser to an AI vendor. All communication is encrypted. We do not use the information you enter to train our models. For the model connections we contract, we also configure the upstream AI vendors not to use it for training. However, this setting does not extend to routes where you have configured your own API key. In that case, whether the data may be used for training is governed by your contract with that vendor.
Credentials for connecting to external systems are stored encrypted with a different key for each company, and that key is itself protected by another key. Because the company and item identifiers are built into the encryption, encrypted data moved to another company cannot be decrypted. No one can view a registered value, and users cannot display it again on screen. The AI model and the code execution environment receive only the item's name, never its value, and the value is actually used in exactly one place, immediately before sending to an external destination. There, we confirm that the destination is allowed, insert the value, send the request, and record the person who ran it, the item used, the destination, and the result in the audit log. In case an external API returns credentials verbatim in its response, we search the response for the values used in that request and replace them with masked text before returning it. This mechanism cannot be disabled.
Sensitive operations require approval before they run. Whether approval is needed is decided by layering our defaults, your settings, and the nature of the conversation. An operation forbidden by any one of these is not executed, even if the others allow it. Operations left unapproved past the time limit end without being executed. For automations, operations that need approval go to the creator of the automation, and if they are not approved within a set time they end without being executed. When generated code is to be run, you review and approve the entire code, not a fragment. Approvals can also be routed by email to approvers, including people outside your company, who do not need an account on the Service. Requiring human approval for committing actions such as finalizing amounts, making payments, and external disclosure is part of the standard design.
Text brought in from outside may contain embedded instructions intended to hijack an agent's behavior. The Service screens knowledge, business data, web search results, the body text of fetched pages, stored files, responses from external integrations, and code execution output before they reach the model. The screening result is kept linked to the text, so if the text is later rewritten, the rewrite itself becomes a clue for detection. By default, the behavior is to detect and record; it can be switched in settings to stop the content from being brought in.
Business operations are recorded in an append-only audit log. There is no means of updating or deleting records, and because each record is chained to the one before it, any rewriting or deletion would be detected. What is recorded is information showing who did what, when, and on what basis, plus values for verification where needed. It does not include the body of knowledge, entered values, credentials, or code. We exclude these because records cannot be deleted from an append-only log, so we would be unable to honor a later request to delete them.
In addition, who used what, when, and how much is recorded in a ledger, and usage can be reviewed daily for any use that breaches your internal rules. The full text of a conversation detected by this review can be viewed only through a path separate from the normal screens, and is shown only when the viewing itself can be recorded. Note that this review is after-the-fact detection; it does not stop a message from being sent.
To improve and maintain the Service, we receive some information outside your tenant. What we receive is conversation titles (a one-line title generated automatically by a model from the content), the user who performed the operation, the names of skills used, ratings and their reasons, and itemized usage charges.
Conversation content never leaves your tenant. We have no function at all, on our side, to open and read an individual conversation. The same applies to comments you write in feedback, the content of imported knowledge and files, and stored credentials.
Viewing the list of titles is recorded in our own audit log, and if it cannot be recorded, the list is not displayed. At present, there is no setting to turn off this provision of information. Its contractual treatment is governed by the Terms of Service and the Privacy Policy.
Regarding storage region, the infrastructure on which the Service runs has a mechanism to pin data storage to a region. However, it covers only the European Union, the United States, and FedRAMP-compliant environments, and it is not possible to specify Japan. The Asia-Pacific region can be requested as a storage location, but under the infrastructure provider's specifications this is best effort, not a guarantee. Since there is also currently no means to confirm the actual storage location, we cannot promise storage within Japan, or that data never leaves Japan. AI inference also uses models from providers outside Japan. Therefore, information you enter may be processed or stored outside Japan. For procurement that requires domestic storage as a mandatory requirement, please treat the Service as not meeting that requirement.
Where information you enter includes personal information, its handling is governed by the Terms of Service and the Privacy Policy. We do not provide it to third parties for our own purposes. If personal information is involved, we will discuss it with you individually at adoption.
Conversations, knowledge, memories, and accumulated procedures are kept until you delete them. Authentication sessions and tokens for external connections are deleted automatically when they expire. Audit logs are append-only and are not deleted. Operational logs are kept for 30 days on a log management service. Storage location, retention periods, and whether secondary use is permitted are set individually at contract.
Kiva is certified under the Privacy Mark and ISMS (ISO/IEC 27001) (ISMS certificate number GIJP-2043-IC; scope: head office). We do not hold a SOC 2 report specific to this service, and we do not claim either ISMAP or LGWAN-ASP.
Adoption starts with reading and drafting only; once accuracy and records have been checked, the scope is widened to executing and registering. We do not take on everything at once.
The OWASP Top 10 for LLM Applications is the OWASP Foundation's list of ten key risks specific to applications that incorporate large language models. This chapter sets out, for each item in the 2026 edition, the measures the Service takes alongside what it does not cover. Related descriptions also appear in "Our Security Measures."
| Item | Measures in the Service | Out of scope |
|---|---|---|
| LLM01Prompt injection | Text brought in from outside passes through a dedicated detection check before it reaches the model. This covers knowledge, business data, web search results, the body text of fetched pages, stored files, responses from external integrations, and code execution output. By default the behavior is to detect and record; a setting switches it to stop the content from being brought in. Changes to that setting are kept in the audit log. Even if instructions are hijacked, writes to external systems do not run without approval. | Text the user enters themselves, and files the user attaches to that conversation, are treated the same as the user's own instructions and are not subject to detection. The structure of pages during browser operation, memories, and recalled past conversations are also currently not subject to detection. |
| LLM02Sensitive information disclosure | Data is held in an independent database per company, and search results are filtered by the requesting user's own permissions. Credential values for external systems are passed neither to the model nor to the code execution environment. If an external response contains credentials, they are replaced with masked text. We do not use the information you enter to train our models. | The no-training setting does not extend to routes where you have configured your own API key. |
| LLM03Excessive agency | Agents have no permissions of their own and act with the permissions of the person who made the request. Operations with real effect, such as writes to external systems, require approval before they run, and those left unapproved past the time limit end without being executed. For automations, approval is requested from the creator, and operations not approved within a set time are not executed. Permissions held on a per-task basis allow only the narrow operations an administrator has defined. | — |
| LLM04Supply chain | Available models are limited to a list we have selected. Only a company's administrators can add connections to external systems. Skills imported from outside are fixed at their content as of import, and bundled scripts do not run until a person has reviewed and approved them. The Service's dependency libraries are built with pinned versions. | There is no mechanism for us to review the external destinations that administrators add. The body of skills imported from outside is not subject to detection. |
| LLM05Data and model poisoning | We have no model of our own and do not train models on your data. When a procedure learned in a conversation is to be kept as a skill, the proposal goes through the detection check, and only those approved by a person are saved. Memories shared across a company or department can be written only by people. Memories phrased as instructions are not saved. | An individual user's personal memories are written by the AI without approval. Memory content is not subject to detection. |
| Item | Measures in the Service | Out of scope |
|---|---|---|
| LLM06Unbounded consumption | Administrators can set spending limits at three levels: the whole company, the default per user, and specific users. Once a limit is reached, further runs stop. A single request, a tool, and a code run each have a time limit. Usage is recorded in a ledger. | There are no per-department limits. Because limits are checked at each break in a run, usage may slightly exceed a limit. There is no setting to limit the number of requests. |
| LLM07Misinformation | Answers based on knowledge or web search show their sources, and sources confirmed by web search are displayed distinctly on screen. Human approval is required before committing actions such as finalizing amounts, making payments, and external disclosure. Adoption starts with reading and drafting only. | There is no mechanism that guarantees the accuracy of AI answers. |
| LLM08Exposure of hidden context | Context passed to the model in a form users cannot see, such as the system prompt and tool definitions, contains no secrets such as credentials. For credentials, the model receives only the name, and the value is inserted immediately before sending to an external destination. Permission checks and approvals take place outside the model and do not depend on this context. So even if it is read out, no secrets leak and no permissions are exceeded. | There is no mechanism to prevent this context itself from being read out. This context includes the memories that the conversation can reference. |
| LLM09Vector and embedding weaknesses | Knowledge search uses full-text search; we do not use a database that stores embedding vectors. Search indexes are kept separately for each company, and search results are filtered by per-user and per-group access rights before they are passed to the AI. By default, knowledge is visible only to the person who registered it. | — |
| LLM10Improper output handling | AI output is not interpreted as HTML; it is displayed on screen as text. HTML and apps created by the AI are displayed inside an area separated from the Service's screens, with no access to sign-in state or other data. Generated code runs in an environment separate for each conversation, and you review and approve the entire code before it runs. | How far approved code communicates externally while running depends on the review at the time of approval. |
What security settings apply to users who access the system environment related to the Service?
For users, an account does not become active until its email address is verified, and at sign-in we always require a one-time code in addition to the password. Users have no way to remove this requirement themselves. You can also set restrictions by source IP address per company. Our employees who can access the Service's production environment are limited to the minimum necessary.
Are general user accounts, privileged user accounts, and system accounts operated with the minimum privileges needed for the work?
Yes. User permissions can be set at three levels, Administrator, Manager, and Member, and narrowed to what the work requires. Agents have no permissions of their own and always act with the permissions of the person who made the request. In addition, the operations the Service can perform on external business systems are limited to the permissions of the accounts you issue to us.
When you outsource work related to the Service, do you conclude non-disclosure agreements (NDAs) with the contractors?
Yes. We select contractors from businesses recognized as handling personal information appropriately. Our outsourcing contracts properly set out security management, confidentiality, and other matters concerning the handling of personal information, and we exercise necessary and appropriate supervision.
What security controls apply to changes to the production environment related to the Service?
Every change to the Service is recorded, and what changed, who made the change, and when can be traced. Only what is actually running in the staging environment can be deployed to production, so only changes that have gone through staging reach production.
Do you define and use processes, procedures, and tools to ensure that servers related to the Service are protected from viruses and malware?
The Service's application is configured so that its entire runtime environment is replaced at every deployment, and we do not make manual changes to running environments. Maintenance of the underlying infrastructure falls within the cloud infrastructure provider's scope of responsibility.
Are security measures in place to prevent, detect, and monitor tampering with content in the server environment related to the Service?
Everything deployed to production is under version control, so you can trace which version is running at any point. In addition, business operations are recorded in an append-only audit log in which each record is chained to the one before it, so any later rewriting or deletion would be detected.
Is data handled on the Service, or sent to related servers, limited to the minimum needed for its functions, and can users be aware of its content?
The Service handles information you enter or upload and information generated from it, information that may help identify users or companies, and statistical, performance, and fault-analysis information for service quality and operation. What an agent sends externally is presented to the user as an approval request before it runs.
Is data sent to servers related to the Service appropriately protected, for example by encryption?
All communication is encrypted, and stored data is encrypted as well. In addition, credentials for external systems are encrypted with a different key for each company, and that key is itself protected by another key.
Is data handled by the Service separated from other companies' data?
Yes. The place where data is stored is itself divided by customer company, and conversation records, business data, accumulated procedures, memories, and audit logs are all held in an independent database for each customer company. Your data never sits alongside other companies' data in a shared database.
Is customer data used to train AI models?
We do not use your information to train our models. For the model connections we contract, we also configure the upstream AI vendors not to use it for training. However, this setting does not extend to routes where you have configured your own API key. In that case, whether the data may be used for training is governed by your contract with that vendor.
By what route is customer data sent to AI vendors?
We send the information needed for processing through routes we manage. Nothing is sent directly from your browser to an AI vendor. All communication is encrypted.
Could the AI operate external systems on its own?
Sensitive operations require approval before they run and are not executed until approved. Operations left unapproved past the time limit end without being executed. For automations, operations that need approval go to the creator of the automation, and if they are not approved within a set time they end without being executed.
Is data stored in Japan?
We cannot promise that. The infrastructure on which the Service runs has a mechanism to pin the storage region. However, it covers only the European Union, the United States, and FedRAMP-compliant environments, and it is not possible to specify Japan. The Asia-Pacific region can be requested as a storage location, but under the infrastructure provider's specifications this is best effort, not a guarantee. There is also currently no means to confirm the actual storage location. In addition, AI inference uses models from providers outside Japan, so information you enter may be processed outside Japan. For procurement that requires domestic storage as a mandatory requirement, please treat the Service as not meeting that requirement.
Does your staff ever view the content of our conversations?
We never view the content of conversations. We have no function at all, on our side, to open and read an individual conversation. What we receive is limited to conversation titles, the user who performed the operation, the names of skills used, ratings and their reasons, and itemized usage charges. A conversation title is a single line generated automatically by a model from the content, and at present there is no setting to turn off this provision of information. Comments you write in feedback, the content of imported knowledge and files, and stored credentials also never leave your tenant. Viewing the list of titles is recorded in our own audit log, and if it cannot be recorded, the list is not displayed.
Do you hold third-party certifications?
Kiva is certified under the Privacy Mark and ISMS (ISO/IEC 27001) (ISMS certificate number GIJP-2043-IC; scope: head office). We do not hold a SOC 2 report specific to this service, and we do not claim either ISMAP or LGWAN-ASP.
What about vulnerability assessments, availability, and incident response?
We answer questions about our operational structure individually in your own security checklist. At adoption, these are finalized in discussion with your IT and audit departments, together with the connection method, how credentials are stored, and log retention periods.
Contact for inquiries about this document
Kiva Corporation, Information Desk
+SHIFT TSUKIJI 7F, 3-12-5 Tsukiji, Chuo-ku, Tokyo, Japan
Tel. +81-3-6264-0825 / info@kiva.co.jp
Inquiries while using the Service are handled through the help channels inside the application. For how-to questions, use the FAQ at the bottom right of the screen; for problems, use "Report a problem" just above it, or send feedback.