Guardrails

Availability

Open core · self-host + all Splyntra Cloud plans

Detection observes and scores risk after the fact. The guard acts on it in real time. Configured on the SDK, the guard runs a pre-flight check on each call and can monitor or block risky input before your agent proceeds.

Guard modes

Set guard on the SDK constructor. It takes three values:

Mode	Behavior
"off"	No inline guarding. Detection and redaction still run.
"monitor"	Risky calls are flagged and scored but allowed to proceed.
"block"	High-precision injection is blocked pre-flight, raising SplyntraBlocked.

Python

from splyntra import Splyntra, SplyntraBlocked

Splyntra(
    api_key=...,
    project="support-agent",
    guard="block",          # "off" | "monitor" | "block"
    guard_fail_open=True,   # default; allow if the guard service is unreachable
)

try:
    run_agent(user_input)
except SplyntraBlocked:
    # A high-precision injection was detected pre-flight.
    reject_request()

TypeScript

import { Splyntra, SplyntraBlocked } from "@splyntra/sdk";

new Splyntra({
  apiKey: ...,
  project: "support-agent",
  guard: "block",         // "off" | "monitor" | "block"
  guardFailOpen: true,    // default; allow if the guard service is unreachable
});

try {
  await runAgent(userInput);
} catch (e) {
  if (e instanceof SplyntraBlocked) {
    // A high-precision injection was detected pre-flight.
    rejectRequest();
  }
}

Pre-flight injection blocking

In block mode the guard performs a high-precision prompt-injection check before the call runs. When it matches, the SDK raises SplyntraBlocked and the call never executes.

Only high-precision injection signatures block. This keeps the guard from tripping on legitimate input — benign role-play and ordinary prompts that resemble instructions pass through. Lower-confidence signals are still scored and surfaced via detection and the risk score, but they do not block.

Fail-open behavior

guard_fail_open / guardFailOpen controls what happens when the guard service cannot be reached. With the default true, calls are allowed if the guard is unreachable, so a guard outage never takes your agent down. Set it to false to fail closed and block when the guard cannot render a verdict.

Next steps

• Security overview — the detection pillars and risk score.
• Detection & redaction — what the detectors flag.