AI Security Self-Assessment

Do you maintain an inventory of every AI/LLM system, its purpose, and its risk classification?

You cannot govern what you cannot see. An inventory is the basis of every framework.

Is there a named owner accountable for the security and risk of each AI system?

Clear accountability ensures risks are actually managed rather than assumed away.

Do you have an acceptable-use / AI policy covering approved models, data, and prohibited uses?

Policy turns intent into enforceable rules for the whole team.

Do you minimize and redact sensitive data (PII, secrets) before sending it to a model?

Data sent to an LLM can be logged, cached, or used for training depending on the provider.

Do you control retention and know whether your provider trains on your prompts?

Retention and training settings determine where your data ends up and for how long.

Where you process personal data, do you have a lawful basis and honor data-subject rights?

Processing personal data through AI still falls under privacy law (e.g. GDPR).

Do you verify the provenance and integrity of models, datasets, and dependencies you use?

Poisoned models/datasets and compromised packages are a growing supply-chain risk.

Do you test prompts and models against adversarial inputs before release?

Red-teaming surfaces jailbreaks and prompt-injection weaknesses before attackers do.

Are system prompts free of embedded secrets, and are credentials injected at the tool layer?

Anyone who extracts your system prompt gets any secret embedded in it.

Do you treat user and retrieved content as untrusted data and defend against prompt injection?

Prompt injection (OWASP LLM01) is the most common and impactful LLM attack.

Do you validate and safely handle model output (no blind execution/rendering)?

Insecure output handling (OWASP LLM05) turns model text into XSS, SSRF, or code execution.

Are agent/tool actions constrained by least privilege and human approval for risky operations?

Excessive agency (OWASP LLM06) lets a manipulated model take damaging actions.

Do you enforce rate limits and cost/usage caps to prevent abuse and denial-of-wallet?

Unbounded consumption (OWASP LLM10) drives cost and outages.

Do you log AI interactions (with sensitive data redacted) for audit and investigation?

You cannot detect or investigate incidents you never recorded.

Do you have an incident-response plan that explicitly covers AI failures and misuse?

AI incidents (harmful output, data leaks, jailbreaks) need a rehearsed response.