How it works

From a key to a fix

No agents to install. You bring the targets, we bring the attacks and the scoring.

  1. 01

    Connect a target

    A model key, an AI-backed endpoint, or a Google or Microsoft mailbox an assistant reads. Keys and tokens are encrypted at rest, and we only use them to run the test you asked for.

  2. 02

    Run a check

    We send synthetic probes: enumeration for keys, injection for endpoints, canary emails for inboxes. Nothing destructive. Each case is judged by signal: a reflected canary, a disclosure signature, a reachable resource.

  3. 03

    Read the findings

    Each case is a card: what was tried, the pattern at altitude, the outcome (caught, partial, escaped), what leaked, and the fix. The check rolls up to a resilience score.

  4. 04

    Fix and re-run

    Assign findings to your security lead, close the gaps, and run it again. Schedule re-tests so a model or prompt change cannot quietly reopen a hole.

What a check covers

Three doors into your AI, one surface to test them

Keys, endpoints and inboxes are the three ways data gets from the outside world to the model and back out again. We probe each the way an attacker would.

Model keyswhat a leaked key reachesAPI endpointswhat your endpoint returnsAgent inboxeswhat an assistant will act onYour AImodel · agent · assistantYour dataFiles & datasetsCustomer recordsSystem promptsKeys & secretsModelInjecttests here

The injection we test for

What a probe actually does

An injection hides an instruction where your AI will read it as if you wrote it. Our probe carries a canary instead of a payload, so a leak is unambiguous and nothing real is at risk.

  1. 1

    Attacker hides an instruction in content your AI reads

  2. 2

    The model reads it as if it came from you

  3. 3

    It plans an action: send, disclose, enumerate

  4. 4

    Data leaves, or Model Inject catches it first in a test

The output

One number you can move

Every check rolls up to a resilience score from 0 to 100, weighted by severity. It is per target and per run, so it climbs when you fix something and drops when a model update reopens a hole.

Caughtdeclined or unreachable; nothing came back
Partiala fragment of the attacker's intent landed
Escapedit worked; data or an action came back
31of 100

Support bot key

Key Exposure check, latest run

1 caught, 1 partial, 4 escaped

Credits meter our side, not yours

A credit is one test case run by Model Inject. Your plan includes a monthly pool, and you top up with packs or auto-recharge. The model tokens a probe burns are billed to you by your provider. We never touch your spend.

Safe by construction

Email tests deliver benign canary messages to a mailbox you connect. API probes run through your real endpoint but ask only for proof, not for harm. Key tests enumerate and ping, and they never delete or modify. You are testing systems you own.

Try it on one target

Bring one key, one endpoint, or one inbox and we run the check live.