Judge an action before it happens

Send the tool call the agent was about to make. Returns whether it may proceed, and when the answer is redact, the same arguments with the sensitive values swapped for stand-ins.

Charged as one request against the same allowance as /redact.

post/inspect

Request

  • Base URL: https://secureai.one/v1
  • URL: https://secureai.one/v1/inspect
  • Auth: HTTP bearer

Request body

toolstring required

The action about to be taken.

inputunknown required
direction'outbound' | 'inbound'
agentstring

Names this agent or run so its actions group together in the trail.

approvalIdstring

Set when coming back after a person has decided. The action's shape is re-checked, so a yes cannot be spent on a different action.

Example request

{
  "tool": "http.post",
  "input": {
    "url": "https://vendor.example.com/send",
    "body": {
      "to": "ana@example.org"
    }
  }
}

Response

A decision.

decision'allow' | 'redact' | 'approve' | 'block' required

The most severe decision across every finding. "approve" means a person must decide before this can happen.

inputunknown
mapobject required

{standIn: real}. The only way to restore the reply; not stored here.

toolDeniedboolean required

True when the tool itself was refused, in which case nothing was inspected.

policySource'account' | 'default' | 'unreadable' required
auditIdstring required
approvalIdstring

Set when the decision is "approve". Poll /approvals/{id}, then re-submit the same input with this id.

expiresAtinteger

When waiting stops being worth it. Milliseconds since epoch.

Example response

{
  "findings": [
    {
      "path": "body.customer.email"
    }
  ]
}

Changes

No recorded changes to this operation across all 1 revision of this API.