AI Security Readiness Quiz

When employees use third-party LLMs, what types of data are they actually allowed to paste in?

There is a written, enforced policy with examples and a quick-reference card.

There is a written policy but enforcement is inconsistent.

There is informal guidance but nothing written.

There is no policy.

How well does your team understand prompt-injection and indirect-prompt-injection risks?

Engineering and security have walked through threat models on real prompts and untrusted inputs.

Senior engineers understand it conceptually but it has not been operationalized.

A few people have heard of it.

It has not been discussed.

How visible is shadow AI usage in your environment?

You have an active inventory of approved tools, with detection for unapproved tools.

You have an inventory of approved tools.

You ask people to self-report.

You assume people are not using AI tools and have not checked.

For decisions informed by AI, what is the human-review posture?

High-stakes decisions require named human approvers and reviewers are trained on AI failure modes.

There is a review step for some decisions but training is inconsistent.

There is review only when something goes wrong.

There is no formal human-in-the-loop.

When AI behaves unexpectedly in production, who gets paged and what is the runbook?

There is a named on-call, a runbook covering AI-specific failure modes, and tabletop exercises within the last 6 months.

There is a named on-call but no AI-specific runbook.

There is generic incident response that would eventually surface AI issues.

There is no defined escalation path.

How do you validate AI outputs that flow back into your systems or to customers?

Schema, type, and content validation, plus monitoring for hallucination/injection signals.

Schema and type validation only.

Spot-check sampling.

You do not validate.

How do you assess AI vendors?

You have a defined assessment covering data residency, training-data use, model lineage, security controls, and exit strategy.

You inherit vendor risk from your standard SaaS process.

You evaluate ad hoc.

You have not formally evaluated AI vendors.

Who owns AI governance at your organization?

A named accountable owner with cross-functional authority and quarterly reporting.

A working group with rotating ownership.

It is unclear who owns it; people defer to security or legal.

No one.

How are AI prompts and responses logged?

Centralized, queryable logs with retention aligned to your data classification, plus PII redaction.

Logged centrally but PII handling is unclear.

Logged inside individual tools only.

Not logged.

When you upgrade or swap an AI model in production, what controls fire?

Versioned model registry, regression evals, staged rollout, and rollback plan.

Manual regression check before swap.

You swap and watch.

You have not yet had to swap a model in production.