Customer Data and Support Boundary¶
This page defines the default self-hosted trust boundary for agent-bom.
The short version:
- customer findings, runtime events, audit, fleet, graph, and tenant data stay in customer-controlled infrastructure by default
- customer operators can review their own tenant-scoped logs, audit, and events through the UI and API when their role allows it
agent-bommaintainers do not get silent access to customer content in self-hosted deployments- outbound telemetry and support sharing are explicit operator choices, not hidden defaults
What customer users can see¶
For an authenticated user inside the customer tenant, the normal control-plane experience can include:
- audit trail and auth/debug history
- scan jobs and scan results
- findings, graph, remediation, and compliance state
- fleet inventory and MCP provenance
- proxy and gateway audit, alerts, and runtime health
- traces and pushed runtime results where those paths are enabled
That access is still role-scoped:
viewerstays read-onlyanalystcan run and review operational workflowsadmincan manage keys, policies, and other protected control-plane actions
The UI may hide or disable actions for clarity, but the API remains the source of truth for what each role is allowed to read or mutate.
What stays customer-owned by default¶
In the self-hosted model, these data classes stay in the customer's own control plane and storage tier unless the operator explicitly exports them elsewhere:
- findings and remediation state
- runtime events and runtime audit
- fleet inventory and endpoint posture
- graph nodes, edges, and correlated evidence
- tenant-scoped auth and audit history
- generated reports and compliance exports
Credential and PII evidence boundaries¶
agent-bom distinguishes between finding a risk and collecting the sensitive
value that caused it.
For MCP and agent configuration inventory, the scanner records credential-like
environment variable names such as OPENAI_API_KEY or DATABASE_URL. It does
not read the environment variable value.
For an explicit project secret scan, the scanner reads files inside the chosen
scope so it can classify likely hardcoded credentials or PII. Findings do not
store the matched value or a prefix of the matched value. They keep only the
relative file path, line number, finding type, severity, and a redacted evidence
label such as [CREDENTIAL_REDACTED].
agent-bom does not use discovered credentials to call providers, validate
whether a token is live, or enrich findings. Any live-secret validation should
be a separate operator-approved workflow with its own network, audit, and
retention boundary.
That is true whether the customer keeps the data in:
- Postgres
- ClickHouse
- S3
- Snowflake
- OTEL / SIEM destinations they control
What agent-bom maintainers do not see by default¶
In self-hosted deployments, agent-bom maintainers should not see customer
tenant content by default.
That includes:
- findings payloads
- prompts or tool arguments
- audit contents
- fleet inventories
- tenant business data
- credential-backed configuration details
Self-hosted means the customer operates the control plane and owns the data plane. There is no silent vendor backhaul.
Minimal product-health telemetry¶
If a customer later wants to share narrow product-health telemetry, the safest default is a minimal, opt-in set such as:
- auth success or failure counts
- error classes and crash reports
- request IDs and trace IDs
- anonymized session IDs
- product version and deployment mode
- component health and route-level latency classes
This is product-health telemetry, not customer content telemetry.
Self-hosted boundary vs provider operations¶
This page describes the self-hosted default:
- customer operators use the product inside their own infra
- customer tenants see their own data by role
agent-bommaintainers do not silently inherit access
That is intentionally different from a future provider-style or MSSP operating model where one operator plane serves many customer organizations. If that product track expands later, it should still preserve explicit access, auditable support flows, and least-privilege defaults instead of weakening the self-hosted trust boundary described here.
Optional exports and interoperability¶
agent-bom is designed so the customer can send their own data to the systems
they already trust.
Common export paths include:
- OTEL collectors
- Splunk
- Elastic
- Datadog
- Snowflake
- ClickHouse
- customer-managed archive and evidence stores
Those exports are customer-controlled integrations. They are not hidden vendor telemetry.
Support access and support bundles¶
The safe support model is:
- support access is explicit
- support sharing is revocable
- support access is auditable
- support data is customer-selected
Today the product already supports a redaction-friendly support bundle flow in the UI help surface. That gives operators a copyable bundle for debugging and bug reports without sending hidden telemetry automatically.
Operators can also inspect the code-generated trust contract through
agent-bom trust --format json and GET /v1/auth/policy. Both surfaces expose
the same credential, network, storage, auth, SCIM tenant, and export
boundaries. The same contract covers connector, plugin, skill, and role
boundaries: agentless read-only connectors by default, operator-scoped
plugin/skill execution, and least-privilege RBAC for every control-plane action.
If a customer needs deeper support later, the recommended model is:
- customer-generated support bundle
- explicit export to the destination they choose
- time-bounded break-glass access only when the operator approves it
One-sentence policy¶
Customers own their findings, audit, runtime events, fleet data, and exports; agent-bom owners should only see minimal product-health telemetry when the customer explicitly opts in or shares support data.