Skip to content
Methodology

The AI Exposure Score

One grade, from A to F, for how much of your AI and agent infrastructure an attacker can see and reach from the open internet — with no login and no active testing. This page is the rubric; there is no black box.

How a scan works

  1. 1Attribute the domain to its internet-facing hosts (certificate transparency, passive DNS, subdomain enumeration).
  2. 2Probe candidate surfaces for AI/agent fingerprints — MCP, A2A agents, model APIs, orchestrators, vector databases.
  3. 3For each confirmed surface, test whether it is authenticated, then read (never call) its advertised capabilities.
  4. 4Grade exposure and exploitability from those passive signals, and assemble the register.

Signals that move the grade

Exposed vs. secured

Every discovered endpoint is tested for an auth wall on a sensitive path. A 401/403 means secured; a service that answers without a credential is exposed. Only exposed surfaces raise your grade.

Exploitability (passive)

For an exposed MCP we read — never call — its tool manifest and check the auth signals the spec defines: OAuth protected-resource metadata, WWW-Authenticate, per-tool auth parameters, dynamic client registration, and PKCE. Destructive or data-access tools with no gating grade higher than read-only ones.

Reachability

CORS posture (an arbitrary-origin allow means browser/DNS-rebinding reachable) and self-signed/expired TLS are treated as amplifiers, not just cosmetic.

Corroboration

Open-source detection templates and read-only MCP posture analysis (YARA) corroborate findings. All detection-only — no attack traffic is ever sent.

The grades

A

No exposed AI/agent surfaces attributable to the domain (authenticated servers are fine).

B

Only low-severity exposure — informational surfaces reachable, nothing dangerous.

C

Medium-severity exposure present — surfaces that warrant review.

D

High-severity exposure — an unauthenticated agent/model surface with reachable capability.

F

Critical exposure — e.g. an unauthenticated MCP whose destructive or data-access tools are callable.

The grade is capped whenever a critical or high-severity exposure is present, so a single unauthenticated, capable MCP cannot be averaged away by otherwise-clean infrastructure.

What stays private until you verify ownership

The public report shows counts and your grade — never the coordinates. Endpoint URLs, exposed tool names, MCP catalogs, and per-asset findings are revealed only after you sign up and prove you own the domain. That way a scan can’t be used as a target list against someone else.

Check your domain →