Authorized traffic
Only send agent traffic, credentials, tools, web targets, and customer data that the user is authorized to use and process.
// draft acceptable use ยท reviewed 2026-07-20
Keep agent execution authorized and governable.
Punk is designed to observe and optimize agent work while keeping side effects attributable and policy-governed. This draft describes sensible usage boundaries; it is not a binding policy or legal advice.
Draft status
These are proposed usage expectations only. A binding acceptable-use policy, enforcement process, and customer-specific obligations require an executed agreement or approved policy.
// intended use
Use the runtime with authority and context.
Accurate context
Set tenant, app, agent, subject, tool, and side-effect metadata accurately so policies, cache boundaries, audit records, and approvals can work as designed.
Deliberate rollout
Start in observe mode for consequential workloads. Keep high-impact or user-visible tool actions approval-required or denied until policy and evidence support expansion.
// conduct to avoid
Do not bypass safety or use the service to cause harm.
Security abuse
Do not bypass authentication, tenant boundaries, policy controls, rate limits, or side-effect approvals; probe private systems without authorization; or conceal a compromise.
Unauthorized actions
Do not use the runtime to perform unlawful, fraudulent, infringing, deceptive, or privacy-invasive activity, including actions the user is not authorized to take.
Unsafe automation
Do not use unreviewed tools or prompts to trigger high-impact actions without appropriate human review, policy, idempotency, and rollback planning.
Secret exposure
Do not put secrets in prompts, workflow inputs, trace-visible metadata, or customer-facing route explanations. Use tenant credentials and supported credential references instead.
Service interference
Do not disrupt service operation, evade technical controls, or overload shared systems beyond agreed limits.
Public disclosure
Do not publish suspected security issues that could expose credentials, tenant data, or a live exploit path. Use the private reporting channel below.
Report a vulnerability privately.
For suspected credential exposure, auth bypass, cross-tenant access, SSRF, policy bypass, unsafe prompt or tool injection, or supply-chain compromise, email the security team rather than opening a public issue.