Skip to main content

Detection & redaction

Availability

Open core · self-host + all Splyntra Cloud plans

Splyntra inspects the free-form text on every span — prompts, model outputs, tool arguments, and retrieved documents — and both redacts sensitive values and scores the risk they represent. This runs on every plan and every edition.

Detectors​

DetectorWhat it flags
Secret detectionAPI keys, tokens, and credentials such as AWS and Stripe keys.
PII detectionPersonally identifiable information (names, emails, and similar).
Content moderationUnsafe or disallowed content.
Tool-guardUnsafe or unexpected tool calls.
Prompt-injection (beta)Attempts to override the agent's instructions.

Each detector contributes to the run's risk score and pins its findings to the spans where they occurred.

Redact by default​

Redaction happens in two layers, so sensitive data is scrubbed before it ever leaves your process and again at the collector:

  1. Client-side, in the SDK. With redact_by_default / redactByDefault enabled (the default, true), the SDK redacts detected secrets and PII from span text before it is exported. Sensitive values never leave your process in the clear.
  2. Collector second pass. The collector runs the detectors again on ingest, catching anything the client did not, and scores the risk.
from splyntra import Splyntra

Splyntra(
    api_key=...,
    project="support-agent",
    redact_by_default=True,   # default; scrub secrets/PII before export
)
new Splyntra({
  apiKey: ...,
  project: "support-agent",
  redactByDefault: true, // default; scrub secrets/PII before export
});
Turning redaction off

Setting redact_by_default=False / redactByDefault: false sends raw span text to the collector so the detectors can see (and flag) the sensitive values in place. Use it only in trusted, non-production settings such as the security demo below.

The security demo​

The security_demo.py example plants AWS and Stripe keys, PII, and a prompt-injection payload into an agent run with redact_by_default=False, so the detectors flag them and the run surfaces CRITICAL/HIGH findings. It is a quick way to see detection and the risk score end to end.

Next steps​