AxiLayer AI · Labs · Free tool

The 3-minute Agent Exposure Check

Twelve questions about the AI agents you run, the tools they call, who approves them and what stops them. You get a band, a score out of 35 and three first steps. Nothing about your systems is collected unless you ask for the 48-Hour Readiness Report at the end.

Start the check See a sample report

Last verified: 30 October 2026

The check

Twelve questions, one honest answer each

Answer for the organization as it is today, not as the policy says it should be. The score is calculated on this page.

01 · Agents in productionHow many AI agents (software that calls tools or takes actions on its own) run in your organization today?

None that we know of A few, in pilots Several, in production Many, across teams

02 · InventoryDo you have one list of every AI agent, the tools and MCP servers it can call, and who owns it?

No list Partial, kept by hand Complete, kept by hand Complete and discovered automatically

03 · Tool approvalWhen an agent gains a new tool or MCP server, who approves it before first use?

Nobody, the developer adds it The team lead, informally A named approver, recorded A named approver, and unapproved tools are blocked

04 · Agent identityDoes each agent act under its own identity and role scope, or do agents share service accounts?

Shared accounts Some agents have their own Every agent has its own identity Own identity, role scope and credential reference on every action

05 · CredentialsHow are the keys and tokens agents use issued and rotated?

Long-lived, rarely rotated Rotated on a schedule Short-lived, scoped per agent Short-lived, scoped, and an inventory flags shared or old ones

06 · High-impact actionsCan an agent move money, change a record of care, change production data or send to customers without a person confirming?

Yes, today Only some actions are gated All high-impact actions wait for a person They wait, and every confirmation is recorded with a reason

07 · Kill switchIf an agent misbehaves right now, how do you stop it?

Find the developer Revoke its account by hand A documented stop procedure One confirmed stop, enforced at the gateway, with a record

08 · Behavior monitoringWould you notice an agent running at ten times its normal call volume or calling tools it never used before?

No Perhaps, from infrastructure alerts Yes, from logs after the fact Yes, in real time, with a stop recommended

09 · LoggingAre tool calls, credential use, network destinations and decisions logged in one tamper-evident record?

Not centrally Logged, not tamper-evident Logged and tamper-evident Logged, tamper-evident and signed

10 · Data reachDo you know which data each agent can read, and in which regions?

No For some agents For all agents For all agents, and the scope is enforced at the gateway

11 · Frameworks and incidentsWhich frameworks apply to your agents, and do incidents carry a reporting clock (for example EU AI Act Article 73)?

We have not mapped them Mapped, no incident clock Mapped, incident clock kept by hand Mapped, clock automatic, evidence re-verifiable

12 · OwnerWho is accountable for every agent, by name?

Nobody yet The developer of each agent A named business owner per agent A named owner, and the board sees the inventory

See where your agents stand No account. Nothing is sent until you choose to.

Your result

0 / 35

ExposedPartly coveredCovered

Get the 48-Hour Readiness Report. AxiReady runs against your estate, read-only, and a qualified AxiLayer AI auditor returns the inventory, the readiness score per agent and the first controls to switch on, within two business days of access.

Your answers travel with the request so the report starts from them. Source: exposure-check. Only contactus@axilayerai.com ever writes back.

Why these twelve questions

The numbers behind the check, 2026

Each question maps to a finding from this year's surveys, breach reports and incidents.

82%

of enterprises found AI agents running in their infrastructure that nobody had registered; 41% found them more than once. Cloud Security Alliance survey of 418 IT and security professionals, fieldwork January 2026, published 21 April 2026.

65%

had an AI agent-related incident in the past twelve months: 61% data exposure, 43% operational disruption, 35% financial loss. Only 21% have a formal process to decommission an agent. Same survey.

11%

automatically block an agent that acts outside its scope; 38% require human approval and 24% only log it. Same survey. AxiSentinel does all three: blocks at the gateway, holds for a person, records.

43%

of breaches now involve shadow AI, up from 20% a year earlier; shadow AI incidents averaged USD 5.39 million. IBM Cost of a Data Breach Report 2026.

92%

of AI-related breaches involved no proper AI access controls, and only 40% of organizations apply access controls to AI models and data. AI governance was absent at 68% of breached organizations. IBM 2026.

USD 1M

added to the average breach cost when AI drove the attack; prompt injection breaches averaged USD 5.89 million. Global average breach cost USD 4.99 million, a record. IBM 2026.

100+

organizations notified by a frontier lab on 1 October 2026 after about 700 of its own evaluation agents chained exposed credentials and two zero-days into a public model platform's production systems between May and July 2026.

2,054

of 9,695 public MCP tool servers analysed by Trend Micro in July 2026 had no authentication at all, and 2,259 were exploitable. Tool servers are the new unpatched perimeter.

named agentic risks in the OWASP Top 10 for Agentic Applications 2026, from goal hijack and tool misuse to rogue agents. Every AxiSentinel control and every test case maps to one.

Sources: Cloud Security Alliance and Token Security, State of AI Agent Security, 21 April 2026; IBM Cost of a Data Breach Report 2026; Trend Micro MCP server analysis, July 2026; frontier-lab incident report and notifications, 1 October 2026; OWASP Top 10 for Agentic Applications 2026.

What the band means

Exposed, partly covered, covered

Exposed (0 to 13)

Agents act with no inventory, no gate and no stop. The first phases of the platform, find and gate, change this inside a month.

Partly covered (14 to 25)

Controls exist in people and documents. They need to become automatic, enforced at a gateway, and recorded.

Covered (26 to 35)

The controls are in place. What is missing is a signature someone outside the organization can re-verify.

Three layers, twelve phases

Twelve phases. One platform. One signature.

Phase 01FindDiscover and score

Phase 02FindClassify and plan

Phase 03ControlApprove every tool

Phase 04ControlName every agent

Phase 05ControlGate every action

Phase 06ControlHold for a person

Phase 07ControlWatch behavior

Phase 08ControlQuarantine and stop

Phase 09ControlMonitor to the article

Phase 10ProveTest and re-test

Phase 11ProveSign the evidence

Phase 12ProveVerify, free, forever

AxiReady finds. AxiSentinel gates, holds, quarantines and stops, Always On or On Demand. A qualified AxiLayer AI auditor signs the result.