Collect only what is necessary
We limit collection to the architecture, authorization, APIs, test data, and logs necessary for the agreed diagnostic purpose.
Information Security PolicyHandle only the information
Handle only the information
that is necessary.
Boundary diagnostics may involve system architecture, authorization, APIs, logs, and the structure of sensitive HR information. We agree the purpose and scope in advance and minimize the information used for testing.
Mock or anonymized data is preferred, and customer information is not sent to an external generative AI service without prior approval.
Core Principle
Because the demonstration addresses HR data, it is designed to avoid real employee records.
The customer-facing demonstration uses fictional Employees C and D and a mock HR database. It tests whether a general employee can reach another employee’s salary, performance-review, or job-transfer information through AI.
- No real employee names, salaries, or evaluations are used
- Only the structure of the customer environment is simulated where appropriate
- The agent is tested to ensure it does not route around a denial through another database, data warehouse, file store, or RAG index
Security Commitments
How diagnostic information is handled
If a contract or NDA establishes stricter requirements, those requirements take precedence.
Prefer mock or anonymized data
When the objective can be achieved without production HR, customer, or personal data, fictional or anonymized data is used.
No unapproved external AI submission
Customer information is not entered into an external generative AI service without approval. If such use is required, the service, purpose, and data sent are explained in advance.
Local diagnostic environments are available
Testing may be performed with a local LLM or in a customer-designated environment so that information is not sent to an Internet-based AI service.
Restrict storage and access
Diagnostic information is stored only in agreed locations and access is limited to the diagnostic personnel. Unnecessary cloud synchronization is avoided.
Avoid production credentials
Dedicated test accounts, short-lived credentials, and least privilege are preferred. Credentials are not placed in reports or source repositories.
Preserve evidence without over-retaining content
Requests, decisions, execution status, errors, and timestamps may be logged. Sensitive content such as salary amounts is masked or excluded where possible.
Agree retention and deletion
The retained items and retention period are agreed per engagement. Information is deleted using a method intended to make recovery impractical after the agreed period.
Subcontracting requires prior approval
If a third party is required, the scope and provider are explained and customer approval is obtained in advance.
Prompt reporting of suspected incidents
If loss, leakage, or unauthorized access is suspected, the potential impact is assessed and the customer is informed promptly.
Non-disclosure agreements supported
An NDA may be executed before detailed system information is disclosed. Required business information is provided during contracting.
Limit operational impact
Mock or test environments are preferred. Production tests require agreed targets, timing, stop conditions, and recovery procedures.
Information Handling
Information that may be handled
| Information type | Default handling approach |
|---|---|
| Architecture and destinations | Reviewed only to the extent necessary to define the diagnostic boundary. |
| Roles and identity methods | Dedicated test roles or simulated role information are preferred over named production accounts. |
| Prompts and agent configuration | Used only for the agreed purpose and not sent to external AI without approval. |
| HR, salary, evaluation, and transfer information | Fictional or anonymized data is the default. Any use of real data requires specific agreement. |
| API secrets and credentials | Production secrets are avoided; dedicated short-lived credentials are preferred. |
| Logs and findings | Evidence needed for repeatability is retained while sensitive content is masked or excluded. |
Before Testing
Items agreed before testing begins
01Information received
02Storage and access
03External AI use
04Retention and deletion
05Stop and reporting process
Contact
Contact →You may consult us before disclosing confidential details.
The initial discussion only requires a high-level description of the AI agent, connected systems, and boundaries you want to verify.