Agent traffic
Model request messages, provider responses, token counts, costs, latency, model metadata, and route explanations may be processed to serve and explain a run.
// privacy posture ยท reviewed 2026-07-20
Know what enters the runtime.
Punk can process the data a customer enables through agent traffic, provider calls, tools, connectors, workflows, and the dashboard. This page describes the present technical posture, not a privacy contract or legal advice.
Important limitation
This is a draft public privacy posture, not a customer-specific privacy notice, DPA, or determination of legal roles. A customer should review its own data flows, providers, and applicable obligations.
// data the product can process
Feature use determines data scope.
Execution evidence
Trace events can cover request, model, tool, workflow, web, policy, DLP, tripwire, approval, learning, replay, and shadow behavior.
Tools and web
When enabled by an application, traced SDK tool inputs/results, managed connector inputs/results, web snapshots, and web-session actions can enter the product.
Workspace data
Workflow definitions, conversations, agents, schedules, human users, org memberships, sessions, invites, settings, and audit events may be stored.
Credential data
Credential metadata is stored and secrets can be stored for providers, workflows, MCP, or managed connectors when configured by the tenant.
Provider disclosure
Live providers receive content needed to answer the request. Their own retention, training, DPA, and vendor terms are separate from Punk and must be reviewed per provider.
// controls and limits
Technical controls need deliberate configuration.
| Current posture | What it means | What it does not mean |
|---|---|---|
| Tenant scoping | Organizations scope runs, traces, settings, credentials, and related runtime records. | It is not a claim of database-level row security or dedicated infrastructure. |
| Retention sweep | Default retention is 90 days unless tenant settings or a job payload override it; a sweep deletes old runs, traces, and audit records for the tenant. | It does not itself establish backup retention, legal hold, or a customer-specific deletion commitment. |
| Redaction and DLP | Tenant redaction masks stored trace strings; streaming DLP masks supported sensitive patterns in returned bytes. | They are not a substitute for data minimization or a guarantee that all sensitive data is detected. |
| Credentials | Full API tokens are shown once and stored as hashes; credential secrets use AES-256-GCM when an encryption key is configured. | It does not establish a complete encryption or key-rotation program for every deployment. |
| Learning | Cross-tenant aggregate learning is a tenant setting and should require explicit opt-in. | It does not authorize sharing raw customer data across tenants. |
// customer review
Questions to resolve before production use.
Map the data classes
Identify data in prompts, traces, tool calls, connectors, web snapshots, provider calls, and customer-facing route explanations. Do not place secrets in prompts or trace-visible metadata.
Set the tenant posture
Confirm retention, redaction, streaming DLP, semantic-cache, and cross-tenant-learning settings for the intended workload.
Review external providers
Confirm which live providers and vendors receive content and assess their terms, retention, training, DPA, and subprocessor posture.
Request contractual review if needed
Data-processing, residency, regulated-data, deletion, and audit obligations are not resolved by this page. Use the request path for a scoped review.
Privacy review starts with the actual traffic.
Bring the configured data classes, enabled providers and connectors, retention target, and deployment evidence. We can identify what the product supports and what needs contractual or operational proof.