Security

Built for operational trust.

Paprika is designed for teams running AI agents in production. Reproducibility, transparency, and execution isolation are core to the platform.

Replay isolation

In replay mode, Paprika never calls live LLM providers or executes live tool functions. All outputs are returned from the recorded trace.

Transparent runtime

Every execution step is recorded as a structured event. No hidden state mutations or opaque runtime decisions.

Policy enforcement

Runtime policies halt execution proactively. Step limits, token budgets, and repeat detection prevent runaway behavior before it causes damage.

Mismatch detection

During replay, Paprika validates that the current execution matches the recorded trace. Any divergence raises an explicit ReplayMismatchError.

Trace control

Traces are structured, portable, and under your control. Storage location and retention are configurable. Enterprise deployments can use private storage.

Data handling

Traces may contain prompts or tool payloads. Redaction and retention policies are configurable. Sensitive data handling is documented and continuously improved.

Security principles

These principles guide every design decision in Paprika.

  • Replay mode must never call live tools or LLM providers.
  • Trace format is structured, documented, and versioned.
  • Policy violations halt execution with structured exceptions.
  • All runtime behavior is inspectable via traces and API.
  • No telemetry or external data transmission without explicit opt-in.
  • Enterprise deployments support private, dedicated infrastructure.