Clark X Group / Venture Pre-production -- privileged enforcement disabled by design

AI that reads your security evidence. Signed humans decide what happens next.

PrivaShield ingests real network telemetry, runs deterministic detection engines across it, and lets a self-hosted LLM read that evidence and recommend a response -- but the model never gets to act. Every privileged action stays simulation-only until a signed, human-approved policy says otherwise.

The one rule the system enforces

AI is advisory-only

A local Ollama-backed model reads detection evidence and returns recommendations -- it never gets privileged access itself.

Enforcement is disabled by default

Every response action (block IP, isolate an interface, terminate a session, quarantine a file) runs in simulation only. enforced=false, no privileged data-plane access.

Policy changes are signed, not typed in

Every policy change is an Ed25519-signed envelope requiring two-person approval, with rollback -- not a config edit.

That split isn't a setting to remember to turn on -- it's structural. The AI layer has no path to a privileged action that doesn't pass through a cryptographically signed, two-person approval first.

What's actually running underneath

Real sensors, not a simulation

Suricata and Zeek run as actual passive network sensors, normalized into a canonical SecurityEvent model.

Deterministic detection, tested like code

DLP, identity-anomaly, and ransomware-behavior engines are checked on every change against a versioned synthetic corpus, with CI that fails on a regression.

Tamper-evident audit log

Every action lands in a hash-chained log with its own verification endpoint -- not a plain database table an admin can quietly edit.

Local LLM, your infrastructure

Threat analysis runs through a self-hosted Ollama model. Nothing about the advisory layer requires sending telemetry to a third-party cloud AI vendor.

Five-role RBAC

Viewer, analyst, operator, administrator, auditor -- gated dashboard and API access, with only token digests stored, never raw tokens.

Apache-2.0, self-hosted

Open-source, run on your own infrastructure via Docker Compose -- Postgres, NATS, Ollama, the API, dashboard, WAF, and both sensor stacks.

How it compares

PrivaShield vs. WazuhA mature open-source SIEM with its own local AI vs. a narrower platform built around signed, advisory-only AI governance

Where it stands today

PrivaShield's own README says this plainly: it is not production-ready as a privileged- enforcement control. External identity-provider integration, SBOM/provenance, and backup/ restore qualification are still on the roadmap, and the packaged quickstart ships with authentication disabled by default for local development -- turning on PRIVASHIELD_AUTH_MODE=local is a deliberate step before any non-loopback exposure. It's early, and it says so.

Get updates

Leave your email and we'll let you know as PrivaShield moves toward a production-ready release:

PrivaShield is an independent, open-source (Apache-2.0) Clark X Group venture -- source, architecture docs, and the full security model: github.com/drewc611/Privashield.