Getting started with PoisonZero
The poison0 agent watches your AI agents’ memory and instruction files, scores every change, and reverts poisoned edits automatically. This page shows how that works in one picture, then walks you through protecting your first device in three steps.
How it protects, in one loop
Every time an agent writes to a watched file, the same loop runs on your device - in milliseconds, before the change is trusted.
Agent changes a file
An AI agent writes to a watched memory or instruction file.
Melira engine scores it
It scores a redacted diff of the change on your device, never your whole file.
A verdict is reached
Based on the score and your sensitivity, one of three things happens:
A dangerous change is rolled back automatically.
A borderline change waits for your confirmation.
A safe change passes through untouched.
Fail-closed by design: if the engine can’t run safely, changes are recorded and surfaced rather than trusted.
Every check runs on your device
The inference engine, the scoring, the verdict - all of it runs on your own machine, in both Cloud and Private mode. Your files are never uploaded anywhere to be analysed.
Your AI agent
Writes memory and instruction files as it works.
Watcher (poison0 agent)
A privileged service that reads each change locally and prepares a redacted diff.
Melira engine
Scores the diff on your device and returns a verdict - in both modes. It never opens your files.
The engine is sandboxed: no network, read-only, and it spawns no processes.
Console
Incidents, audit and fleet settings for your whole team.
Only verdicts, signals and telemetry leave your device - never the contents of your files. Scoring runs locally in both Cloud and Private mode.
Cloud or Private - same local checks, different egress
You pick the mode when you install. Either way every check runs on your device; the mode only decides what (if anything) leaves it.
Model and console fully self-hosted.
Console, model updates and fleet management hosted by us.
* One signed licence check-in every 30 days to update.poisonzero.com. Your files and verdicts never leave the device.
Protect your first agent in four steps
-
1
Create your account and pick a mode
Sign up and pick a mode: Cloud (managed console and model updates) or Private (everything on your own infrastructure). Either way, scoring runs on your device.
-
2
Follow the onboarding guide
The onboarding guide walks admins and developers through the rollout in six steps: prepare the console, install, set up, use, view telemetry and give feedback. Also available as a PDF.
-
3
Install and enroll in one command
Grab an enrollment code from the console, then run a single install command that carries it. It installs the privileged service and enrolls the machine in one shot - on Linux, macOS or Windows, so a rogue agent can’t stop it.
-
4
Watch it work
Protection starts immediately. From the console you review what the poison0 agent caught, under Incidents and Audit.
-
+
Optional: check tool calls with exec0
exec0 checks every tool call of your AI agents before it runs and requires a GPU suitable for Melira Ultra. Choose Off, Observe or Protection under Settings → Protection & Devices → Profiles → Edit on the profile → exec0 with Melira AI → Mode, on private devices on the device with
sudo poison0 exec0 observeorenforce; details in the onboarding guide.
What happens next
From here, the poison0 agent watches in the background and reverts poisoned changes on its own. You review what it caught in the console, under Incidents and Audit. Want the deeper picture? These are the best next reads: