Authorization boundaries
Verify that employee, administrator, HR, department, and tenant roles remain effective when actions are performed through AI.
- Self vs. another employee
- Department and tenant boundaries
- Read, update, and delete rights
Scope of DiagnosticsWhat we test,
What we test,
and what we do not.
We do not grade the general intelligence of an AI model. We test whether an AI agent stays within the user’s authorization, business data, permitted actions, tools, and network destinations. The scope is documented and agreed before testing begins.
Core question: who can reach which information, through which route, and perform which action?
Primary Scope
Primary diagnostic areas
We examine the whole system surrounding the agent: identity, authorization, data, APIs, tools, network paths, and audit evidence.
Data boundaries
Check whether the agent moves beyond its approved search range into another database, RAG index, file store, or analytics platform.
- HR, salary, evaluation, transfer data
- Other departments or users
- Cross-session data mixing
Action boundaries
Check whether a read-only user can cause updates, deletion, export, or external transmission through the agent.
- API and database updates
- File creation and deletion
- Email and external transmission
Tool and network boundaries
Verify that the agent does not select an unapproved tool or destination to bypass the intended path.
- MCP and external APIs
- Web, SaaS, and command execution
- Ports, proxies, and destinations
Instruction and decision boundaries
Test ambiguous instructions, privilege escalation requests, and instructions embedded in external content.
- Prompt injection
- Fake urgency or executive orders
- Multi-turn policy evasion
Auditability and repeatability
Confirm that decisions, actions, denials, and failures can be traced and remain stable under repeated testing.
- Request, decision, and execution logs
- No data access after denial
- Fail-closed behavior
HR Database Demo
DEMO / MOCK DATACustomer-facing demonstration using a mock HR database
No real employee records are used. Fictional employees and a mock HR database demonstrate the principle that an AI acting for an employee must not access information the employee cannot access directly.
Employee CMay view own data only
→
AI AgentActs on behalf of the user
→
Authorization GatewayChecks user, subject, and action
→
Mock HR DatabaseSalary, evaluation, transfer data
Employee C requests their own payslip through the approved route.
The request exceeds Employee C’s rights and must be rejected.
The agent must not try a payroll database, data warehouse, file store, or RAG index after denial.
Conditional / Excluded
Conditional scope and standard exclusions
Available under agreed conditions
- Production testing with agreed targets, stop conditions, and recovery procedures
- Real data when necessary and subject to an agreed handling method
- External SaaS or third-party systems with explicit owner authorization
- High-privilege actions using dedicated accounts, approvals, and backups
- Adversarial testing with written limits on method and impact
Normally outside scope
- General model knowledge, writing quality, translation quality, or benchmark scoring
- Final legal opinions on regulatory compliance, copyright, or privacy law
- Unauthorized intrusion, data destruction, credential theft, or denial of service
- A guarantee that the entire product or system is free from vulnerabilities or future incidents
- Full source-code quality review, load testing, or general defect testing
- Formal certification or audit services such as ISMS, SOC 2, or PCI DSS
Scoping Process
How the scope is defined
Before testing, we agree the included systems, excluded areas, data to be used, stop conditions, and deliverables.
01Review architecture
02Map users and roles
03Define boundaries
04Agree inclusions
05Finalize test conditions
Next
Plans & Pricing →Review the available plans and pricing.
Start with a demonstration or proceed to a structured diagnostic engagement.