Solution · Runtime Guardrails

Posture tells you what could break.
Guardrails stop it in flight.

Scans find the vulnerabilities. Guardrails defend the live request path. The Klyvra Inspector sits on the ingress and egress of your LLM: your app calls it for each prompt, completion, tool call, tool result, and retrieved chunk, and gets back a per-guardrail verdict - block, require human review, redact, flag, or allow. It is a decision service, not a proxy: it returns an opinion, your application enforces it.

See a demo →Back to platform

Solution · Runtime Guardrails

A decision service, not a proxy.

A proxy sits in the data path and makes the enforcement call for you - a new single point of failure on every request. The Klyvra Inspector does not. Your application calls POST /v1/inspect with the content it is about to send or return, and gets back a verdict per guardrail. You decide what to do with it. Two evaluation tiers back every guardrail: a Fast tier (in-cluster Go regex, signatures, and allowlists) that returns in single-digit milliseconds, and an Accurate tier (Lelouch LLM judges) for the cases that need real reasoning.

The guardrail catalog

Six families, one inspection call.

Fifteen guardrail templates across six families cover the prompt, the RAG context, the agent, and the data on both sides of the model. Each one runs Fast (in-cluster, single-digit ms) or Accurate (Lelouch LLM judges) - your call, per guardrail.

Prompt & indirect injection.

Detect direct prompt-injection and jailbreak attempts on ingress, and indirect injection smuggled in through retrieved RAG content before it ever reaches the model.

Jailbreak & topical control.

Catch DAN-class jailbreaks, persona shifts, and toxicity, and keep the model inside its lane with a configurable topical filter on what it will and will not discuss.

PII & secret leakage.

Block PII and credentials on the way in and the way out. An essentials secrets pack (31 detectors) and an exhaustive pack (200 gitleaks detectors) catch keys and tokens before they leave.

Agent & tool guardrails.

Inspect tool arguments for injection, enforce tool authorization, and gate sensitive actions behind a verdict - so an agent cannot be coaxed into a privileged call it was never meant to make.

Stateful trajectory checks.

Some attacks only appear across turns. Stateful guardrails watch for excessive agency and multi-step drift over a whole session, not just one isolated request.

Bring-your-own blocklists.

Ingress and egress blocklists backed by reusable, themed word packages. Whole-word match by default, up to 2,000 terms, with match speed independent of list size. Ship empty; you fill them.

Who it's for

Built for the teams who own the live request path.

Runtime guardrails are for the AppSec, platform, and AI-engineering teams who run the production endpoint - the people who get paged when a prompt injection lands or an agent reaches somewhere it should not. Posture scans tell them what could break in testing; guardrails defend the surface a scheduled scan can never reach: the live, in-flight request.

What this unlocks

Outcomes you can defend in a review.

Outcomes Klyvra customers and design partners use to justify the programme to their boards, auditors, and clients.

Defend the surface scans can't reach.

Scheduled scans test a snapshot. Guardrails inspect every real request, every completion, every tool call - the live traffic, as it happens.

Single-digit-millisecond Fast mode.

The Fast tier runs in-cluster as Go regex, signatures, and allowlists - low enough latency to inline on the request path without users feeling it.

One dashboard for both halves.

Offline posture and inline defense live in the same platform. The vulnerabilities a scan finds and the guardrails defending against them are one view, not two tools.

Sovereign and zero-egress.

The Inspector runs entirely inside your cluster. Prompts, completions, and verdicts never leave your network - the same hard requirement that governs the rest of Klyvra.

One platform.
Every way to deploy it.

AI Posture Management

Put a guardrail
on a live endpoint.

Bring an endpoint and a guardrail you care about - prompt injection, PII, an agent tool gate. We will wire up an inspect call during a 30-minute walkthrough and show you the verdict path end to end.

Request a demo →[email protected]

Posture tells you what could break.Guardrails stop it in flight.