Concept — built and unit-tested, live network demo pending
A support agent that can't be talked into anything.
Most chatbot deployments are one model, no filter, and no record of why a reply was allowed through. This one runs every draft reply past real prompt-injection detection before it ever reaches your customer — and signs every response so tampering downstream is detectable, not assumed away. The pipeline is built and covered by an automated test suite; it hasn't completed a live demo call yet.
How it works
Three steps, not one black box.
Built on the same detection logic already shipping inside Leviathan Platform — not a separate, unproven safety layer.
- 1. First contact
- A fast model drafts a reply to the customer's message.
- 2. Real filter
- The draft and the original message run through Leviathan Platform's actual prompt-injection and exfiltration detectors — the same logic protecting every other product in the line.
- 3. Escalation, only if flagged
- A stronger model reviews the flagged draft plus the detector's finding, and either sanitizes or rejects it. A malformed review response fails closed to a fixed decline, never a guess. Clean replies never touch this step.
Trust & release state
What's real, and what isn't yet.
What's real: the gateway's actual code — the three-step pipeline, the fail-closed escalation logic, and the Ed25519 signing (the same canonical-signing recipe used elsewhere in the platform) — is built and covered by an automated test suite exercising every real decision path: clean passthrough, injection escalation with sanitize/reject, a leaky-draft catch on the reply side alone, signature verification, and tamper detection. The request format has also been verified byte-correct against a real local Ollama instance.
What isn't built yet: a completed live demo call. The dev environment's GPU is hitting an intermittent driver-level crash under model load — a known hardware issue, not a flaw in the gateway logic — so a full network round-trip hasn't gone through cleanly yet. There's also no hosted/decoupled deployment, no enterprise integrations (SIEM/SSO), no published false-positive rate, and no independent benchmark.
What won't happen: this page claiming a live demo succeeded before it actually has.
What isn't built yet: a completed live demo call. The dev environment's GPU is hitting an intermittent driver-level crash under model load — a known hardware issue, not a flaw in the gateway logic — so a full network round-trip hasn't gone through cleanly yet. There's also no hosted/decoupled deployment, no enterprise integrations (SIEM/SSO), no published false-positive rate, and no independent benchmark.
What won't happen: this page claiming a live demo succeeded before it actually has.
Get in touch
Building a chatbot that shouldn't be talked into anything?
Tell us what you're building and we'll figure out if this is a fit.
Talk to us →