TDM Agent Security and Privacy

The TDM Agent has no privileges of its own: it runs as the calling user, sends the LLM only TDM metadata, value lists and counts, and can only write a draft execution that TDM must still authorize and a person must start. This article describes what data reaches the LLM, the security controls, retention and audit, and the split of responsibility between K2view and the customer. It is intended for information-security and architecture reviewers.

For the components and the request flow, see TDM Agent Architecture.

Data Sent to the LLM

The LLM sees TDM configuration and selection metadata, plus value lists and counts. It never receives source records: the agent has no SQL or free-query tool, and no tool returns entity rows.

Data category Example Sent to LLM
User request "6 Gold customers from Production to UAT" Yes
TDM task definitions Task name, action, selection method, editable fields Yes
Entity IDs in an Entity List task The task's stored ID list Yes, when such a task is a candidate. This applies when the user's request contains IDs, and is under the user's responsibility.
Business Entities and environments BE names, environment names and descriptions Yes
Business-parameter metadata Parameter names, types, ranges Yes
Business-parameter value examples Distinct values such as tier = Gold, Silver Yes
Entity counts "412 customers match" Yes
Custom Logic flow parameters Parameter names and lookup values Yes
Source-environment entity records (rows) Customer name, address, balance No
Credentials, API keys, connection strings — No

LLM provider and data residency. The agent calls the environment's default LLM interface. The customer chooses and configures the provider:

  • Anthropic
  • AWS Bedrock (with optional Bedrock Guardrails for PII redaction)
  • Azure OpenAI or OpenAI
  • Google Gemini / Vertex AI
  • A self-hosted model through Ollama (preview)

With a private endpoint or a self-hosted model, no data leaves the customer's network. K2view does not host or proxy the LLM traffic. See LLM Connectors.

MCP tools. Results returned to an MCP client go to that client's own model, not to the LLM configured in Fabric. They are the same kinds of lookup results listed in the table above, plus draft and execution results. The customer's policy for that AI client applies.

Reducing exposure further. Value examples come from business-parameter definitions. If parameter values are sensitive (for example, if they hold personal identifiers), mask them in the parameter definition or exclude them from the parameter list.

Security Controls

The agent has no privileges of its own. It runs as the calling user, cannot run queries or Fabric commands, and can only write a draft execution that TDM must still authorize and a person must start.

Area Control
Authentication Standard Fabric authentication: SSO/OAuth, JWT, API key, Fabric session or Basic. See Fabric Web Services Security.
Authorization Fabric RBAC on each web service. Chat and REST API web services run without elevated permissions.
TDM permissions Task search applies the user's TDM permission group. Starting an execution goes through TDM's normal role and environment checks. See Permission Groups (User Types) and Environment Permission Sets Tab for more information about the TDM permissions.
API identity Each test case runs with the submitting user's name and Fabric roles.
Tool surface An explicit tool allow-list per agent. Built-in SQL and Fabric-command tools are disabled. Calls to tools outside the list are blocked.
Write scope One draft execution record per prepared case, marked as AI-created and linked to the agent session, only on Active tasks. Internal-only override keys are rejected.
Execution None by the agent. A user starts the run from the review or execution link.
Guardrail hooks A pre-tool hook blocks an incomplete preparation. Post-tool hooks end the run once preparation completes or an edit fails.
Run limits A capped number of LLM iterations per turn, continuations and sub-agent depth.
Secrets LLM credentials live in the Fabric interface, kept in the Fabric secret manager or provided by cloud identity (IAM role, ADC), and never in the agent or the prompt. See Secrets Management.
Encryption in transit HTTPS/TLS between client and Fabric, and between Fabric and the LLM provider
Encryption at rest Inherited from the Fabric storage and TDM database configuration

Prompt rules (never execute, never reveal override JSON) are a second layer. The hard boundaries are the tool allow-list, the hooks and TDM's own permissions.

MCP access. Fabric exposes each deployed Data Product as an MCP server, and this is enabled by default. MCP clients authenticate with the same Fabric methods as the web services, and Fabric RBAC authorizes them at Data Product and entity-instance level. Permissions cannot be set per tool. Most TDM MCP tools are read-only lookups, and task search returns only tasks in groups the user may see. Two tools act: CreateExecutionDraft saves an execution draft, and StartTaskExecution starts a task execution. A started execution runs with the calling user's identity and passes TDM's permission and execution validations. Because permissions cannot be set per tool, any role with MCP access to the TDM LU can start the executions its TDM permissions allow. See TDM MCP Tools and Data Product MCP.

Retention and Audit Trail

Every conversation is recorded inside the customer's Fabric deployment and deleted by a scheduled cleanup after a configurable window.

Retention

Setting (aifusion LU global) Default Range Applies to
RETENTION_DAYS 30 days 7–365 Closed conversations: session records, traces, snapshots
EVALUATION_STORAGE_RETENTION_DAYS 90 days 30–365 Evaluation snapshot files no longer referenced

Cleanup is a built-in job that the customer turns on and schedules as needed.

Audit trail

What is recorded Where
Each request and answer, with timestamps Session step records
Every LLM message, tool call (arguments and result) and token usage Session trace tables (aifusion LU instance per session)
Who started each conversation, with which agent Session record in the Insights database
Each API request, its test cases and the per-case agent session ID Request status records in the TDM database
Each prepared TDM execution TDM execution records, under TDM's own audit
Runtime errors Fabric server log (the user sees a generic message)

The trace holds what the agent saw. Access to trace endpoints and the Insights database should be limited to operators.

Shared Responsibility

K2view delivers the agent and its built-in controls. The customer owns the deployment, the identities and the choice of LLM.

Area K2view Customer
Agent logic, tool allow-list, hooks Delivers and maintains Deploys the extension
Hosting, network, encryption at rest — Operates Fabric and TDM in its own environment
Identity and roles Fabric RBAC and TDM permission model Maps users to roles and grants the agent web services only to approved roles
LLM provider Connectors for the supported providers Selects the provider, region and endpoint, and holds the provider contract
Sensitive parameter values Masking capabilities in Fabric Decides which business parameters expose values
Retention Configurable cleanup jobs Sets retention windows and schedules cleanup
Running executions The agent never starts them; the MCP StartTaskExecution tool starts them only on a client's request Reviews and starts each execution, and decides which roles and AI clients may start executions through MCP
  1. Use a private or in-region LLM endpoint (for example Bedrock in the customer's AWS account or Azure OpenAI), or a self-hosted model.
  2. Grant the chat and test-data-agent/* web services only to tester and pipeline roles. Limit trace and observation endpoints to operators.
  3. Give each CI pipeline or test-management tool that calls the REST API its own Fabric user or API key, with only the REST API role. Its requests are then traceable to that tool and can be revoked separately.
  4. Turn off Fabric MCP if it is not used. Otherwise, give MCP access to the TDM Data Product only to approved roles and approved AI clients. MCP access includes the StartTaskExecution tool, which runs tasks.
  5. Review which business parameters expose value lists. Mask or remove any that hold personal data.
  6. Set the retention window to match the customer's policy, and turn on the cleanup schedule.

Previous

TDM Agent Security and Privacy

The TDM Agent has no privileges of its own: it runs as the calling user, sends the LLM only TDM metadata, value lists and counts, and can only write a draft execution that TDM must still authorize and a person must start. This article describes what data reaches the LLM, the security controls, retention and audit, and the split of responsibility between K2view and the customer. It is intended for information-security and architecture reviewers.

For the components and the request flow, see TDM Agent Architecture.

Data Sent to the LLM

The LLM sees TDM configuration and selection metadata, plus value lists and counts. It never receives source records: the agent has no SQL or free-query tool, and no tool returns entity rows.

Data category Example Sent to LLM
User request "6 Gold customers from Production to UAT" Yes
TDM task definitions Task name, action, selection method, editable fields Yes
Entity IDs in an Entity List task The task's stored ID list Yes, when such a task is a candidate. This applies when the user's request contains IDs, and is under the user's responsibility.
Business Entities and environments BE names, environment names and descriptions Yes
Business-parameter metadata Parameter names, types, ranges Yes
Business-parameter value examples Distinct values such as tier = Gold, Silver Yes
Entity counts "412 customers match" Yes
Custom Logic flow parameters Parameter names and lookup values Yes
Source-environment entity records (rows) Customer name, address, balance No
Credentials, API keys, connection strings — No

LLM provider and data residency. The agent calls the environment's default LLM interface. The customer chooses and configures the provider:

  • Anthropic
  • AWS Bedrock (with optional Bedrock Guardrails for PII redaction)
  • Azure OpenAI or OpenAI
  • Google Gemini / Vertex AI
  • A self-hosted model through Ollama (preview)

With a private endpoint or a self-hosted model, no data leaves the customer's network. K2view does not host or proxy the LLM traffic. See LLM Connectors.

MCP tools. Results returned to an MCP client go to that client's own model, not to the LLM configured in Fabric. They are the same kinds of lookup results listed in the table above, plus draft and execution results. The customer's policy for that AI client applies.

Reducing exposure further. Value examples come from business-parameter definitions. If parameter values are sensitive (for example, if they hold personal identifiers), mask them in the parameter definition or exclude them from the parameter list.

Security Controls

The agent has no privileges of its own. It runs as the calling user, cannot run queries or Fabric commands, and can only write a draft execution that TDM must still authorize and a person must start.

Area Control
Authentication Standard Fabric authentication: SSO/OAuth, JWT, API key, Fabric session or Basic. See Fabric Web Services Security.
Authorization Fabric RBAC on each web service. Chat and REST API web services run without elevated permissions.
TDM permissions Task search applies the user's TDM permission group. Starting an execution goes through TDM's normal role and environment checks. See Permission Groups (User Types) and Environment Permission Sets Tab for more information about the TDM permissions.
API identity Each test case runs with the submitting user's name and Fabric roles.
Tool surface An explicit tool allow-list per agent. Built-in SQL and Fabric-command tools are disabled. Calls to tools outside the list are blocked.
Write scope One draft execution record per prepared case, marked as AI-created and linked to the agent session, only on Active tasks. Internal-only override keys are rejected.
Execution None by the agent. A user starts the run from the review or execution link.
Guardrail hooks A pre-tool hook blocks an incomplete preparation. Post-tool hooks end the run once preparation completes or an edit fails.
Run limits A capped number of LLM iterations per turn, continuations and sub-agent depth.
Secrets LLM credentials live in the Fabric interface, kept in the Fabric secret manager or provided by cloud identity (IAM role, ADC), and never in the agent or the prompt. See Secrets Management.
Encryption in transit HTTPS/TLS between client and Fabric, and between Fabric and the LLM provider
Encryption at rest Inherited from the Fabric storage and TDM database configuration

Prompt rules (never execute, never reveal override JSON) are a second layer. The hard boundaries are the tool allow-list, the hooks and TDM's own permissions.

MCP access. Fabric exposes each deployed Data Product as an MCP server, and this is enabled by default. MCP clients authenticate with the same Fabric methods as the web services, and Fabric RBAC authorizes them at Data Product and entity-instance level. Permissions cannot be set per tool. Most TDM MCP tools are read-only lookups, and task search returns only tasks in groups the user may see. Two tools act: CreateExecutionDraft saves an execution draft, and StartTaskExecution starts a task execution. A started execution runs with the calling user's identity and passes TDM's permission and execution validations. Because permissions cannot be set per tool, any role with MCP access to the TDM LU can start the executions its TDM permissions allow. See TDM MCP Tools and Data Product MCP.

Retention and Audit Trail

Every conversation is recorded inside the customer's Fabric deployment and deleted by a scheduled cleanup after a configurable window.

Retention

Setting (aifusion LU global) Default Range Applies to
RETENTION_DAYS 30 days 7–365 Closed conversations: session records, traces, snapshots
EVALUATION_STORAGE_RETENTION_DAYS 90 days 30–365 Evaluation snapshot files no longer referenced

Cleanup is a built-in job that the customer turns on and schedules as needed.

Audit trail

What is recorded Where
Each request and answer, with timestamps Session step records
Every LLM message, tool call (arguments and result) and token usage Session trace tables (aifusion LU instance per session)
Who started each conversation, with which agent Session record in the Insights database
Each API request, its test cases and the per-case agent session ID Request status records in the TDM database
Each prepared TDM execution TDM execution records, under TDM's own audit
Runtime errors Fabric server log (the user sees a generic message)

The trace holds what the agent saw. Access to trace endpoints and the Insights database should be limited to operators.

Shared Responsibility

K2view delivers the agent and its built-in controls. The customer owns the deployment, the identities and the choice of LLM.

Area K2view Customer
Agent logic, tool allow-list, hooks Delivers and maintains Deploys the extension
Hosting, network, encryption at rest — Operates Fabric and TDM in its own environment
Identity and roles Fabric RBAC and TDM permission model Maps users to roles and grants the agent web services only to approved roles
LLM provider Connectors for the supported providers Selects the provider, region and endpoint, and holds the provider contract
Sensitive parameter values Masking capabilities in Fabric Decides which business parameters expose values
Retention Configurable cleanup jobs Sets retention windows and schedules cleanup
Running executions The agent never starts them; the MCP StartTaskExecution tool starts them only on a client's request Reviews and starts each execution, and decides which roles and AI clients may start executions through MCP
  1. Use a private or in-region LLM endpoint (for example Bedrock in the customer's AWS account or Azure OpenAI), or a self-hosted model.
  2. Grant the chat and test-data-agent/* web services only to tester and pipeline roles. Limit trace and observation endpoints to operators.
  3. Give each CI pipeline or test-management tool that calls the REST API its own Fabric user or API key, with only the REST API role. Its requests are then traceable to that tool and can be revoked separately.
  4. Turn off Fabric MCP if it is not used. Otherwise, give MCP access to the TDM Data Product only to approved roles and approved AI clients. MCP access includes the StartTaskExecution tool, which runs tasks.
  5. Review which business parameters expose value lists. Mask or remove any that hold personal data.
  6. Set the retention window to match the customer's policy, and turn on the cleanup schedule.

Previous