Audit and attribution

Record the decision where enforcement happens.

Every Safe Boundary surface emits structured decision data: what was allowed, blocked, or masked and the caller information that surface can actually establish.

Get early access
Review the security model →

Why this exists

Useful evidence starts with honest attribution

MCP writes rotating local JSONL and can mirror records to the console on a best-effort basis. The wire path uses a bounded event buffer with optional Parquet output. REST can uplink policy decisions to the same console shape.

Human SSO sessions can carry certificate-verified identity. PostgreSQL roles come from upstream authentication. MCP agent names are operator-supplied attribution. The audit keeps those meanings distinct.

What it does

Concrete controls, stated precisely.

Decision context

Records include outcome, reason, request context, masked columns where applicable, and available subject metadata.

Local-first storage

MCP JSONL and wire-path Parquet options keep inspectable records inside the deployment.

Rollout visibility

Observe mode and counters show which rules would fire before production behavior changes.

Evidence example

A useful event shows the decision and its provenance.

Illustrative audit shape. Storage, retention, and SIEM delivery remain deployment choices.

Sample event

{
  "outcome": "masked",
  "reason": "configured_column",
  "surface": "mcp",
  "masked_columns": ["customers.email"],
  "subject": {"kind": "agent_label", "value": "analyst-tool"}
}

Operator question

What happened?
Which policy reason fired?
Which fields changed?
How strong is the caller identity?

How it works

A visible enforcement sequence.

  1. Capture

    Emit an event at the policy decision point.

  2. Attribute

    Attach verified or supplied identity according to the access surface.

  3. Persist

    Write to the configured local sink and optionally mirror to the console.

  4. Export

    Use JSONL or Parquet in your existing analytics and retention pipeline.

Protected when

  • ✓Policy decisions made by Safe Boundary
  • ✓Local inspection and export workflows
  • ✓Observe-mode rollout analysis

Important boundaries

  • •Console uplink is not the authoritative MCP record
  • •Retention, immutability, alerting, and SIEM routing depend on deployment
  • •Supplied MCP labels are not verified workload identity

Continue the story

Related Safe Boundary pages

Security model →

Understand trust boundaries, credentials, identity, and bypass conditions.

MCP server →

Give AI tools a bounded PostgreSQL query surface with local audit.

Human SSO →

See the path with verified human-session identity.

Test the boundary on a real access path.

Start in observe mode and validate the policy against representative traffic before you enforce it. That gives the team evidence for each decision, along with explicit bypass controls for the cases that need a deliberate exception.

Get early access