Host-based defense against adversarial AI

Demask theattacker.Control thetempo.

Ouija is host-based endpoint defense against adversarial AI. A DPU isolates the host, specialized local models decide without waiting on an API round trip, and Active-Reactive Behavioral Detection™ is built to identify the model behind an attack, however it's disguised.

Closed beta · Technical and government evaluations

Ouija · session traceSimulated
Session
7f3a-91c
Cadence
142 ms ±0.6
Model class
LLM-driven agent
Objective
Credential discovery
Origin
Rented GPU host
Action
Held at boundary

Held at boundary · fail-closed

Illustration with simulated sessions. One session keeps a machine-regular cadence. Ouija flags it, identifies it, and holds it at the boundary while the other traffic keeps moving.

The product

Intelligence at the boundary.

Ouija

Host-based defense against adversarial AI. Now in closed beta.

Machine speed is needed to act against machines. Ouija runs on a DPU in front of the host, so every packet stops there first, and specialized local models decide on the spot.

No LLM operates in the defensive path. Your data stays on your systems, with no outside API dependency announcing your presence.

Explore the closed beta
Where Ouija sits Inbound traffic passes through Ouija on the DPU before it reaches the host. One stream is held at the DPU. An operator sets rules, posture and response. Operator RULES · POSTURE Inbound TRAFFIC Ouija ON THE DPU FAIL-CLOSED Host PROTECTED HELD
Ouija runs on the DPU between the network and the host. Decisions happen locally, and the operator decides what happens next.

Behavioral intelligence

Reads patterns, sequence, and response, so attacks are recognized by behavior instead of exact signatures.

Local decisions

Specialized models on the machine decide on the spot, with no waiting on an API round trip.

Human authority

Operators see why an action fired and choose the rules, response posture, and level of protection.

It began
with fraud.

Behavioral Outlier Optimization was discovered while coding models for fraud discovery. It became a shared primitive for detection, attention, and surprise: noticing change, deciding where to look, and recognizing when expectations need to change.

Explore the primitive

The method

The evidence is
part of the work.

Every model is trained and challenged with live activity on infrastructure we own and control, so it keeps the shape, sequence, and timing that static datasets lose.

  1. 01

    Observe

    Capture behavior with its context.

  2. 02

    Challenge

    Test under changing conditions.

  3. 03

    Refine

    Keep the result.
    Improve the system.

Inside Creator Fusion

Research notebook

Questions worth pursuing.

  1. 01

    Active–Reactive Deception

    What a response can reveal.

  2. 02

    Observer Masking

    What changes when a system knows it is watched.

  3. 03

    The One-Bit Alarm

    When a small observation carries weight.

  4. 04

    The Cost of Remaining Hidden

    Change the game. Make concealment cost something.

  5. 05

    Masking Capacity and Residual Detectability

    How much room does a task leave for camouflage?

  6. 06

    Hydrodynamic Fisher Routing

    A shared geometry for where information goes next.

Full papers are available by approved request.

Explore research themes

Bring the hard problem.

Technical evaluations, research collaboration, and problems that don't fit a category.

Request an evaluation