When a business ships an LLM-powered feature (chatbot, summariser, agent), the security risks are not the same as for a traditional web app. OWASP publishes a Top 10 list specifically for LLM applications. It is free, community-maintained, and a good starting checklist.

The risks that matter most in practice

OWASP LLM riskWhat it means in business termsTypical first control
Prompt injectionA user tricks the model into ignoring your instructions or leaking dataSeparate trusted instructions from untrusted input; validate outputs
Sensitive information disclosureThe model reveals PII, secrets, or training dataRedact before sending; scope retrieval; log and audit
Excessive agencyAn agent takes actions it should not (emails, payments, deletions)Least-privilege tools; human approval for irreversible actions
OverrelianceUsers trust model output that is wrong or fabricatedShow sources; require verification for high-stakes outputs
Insecure output handlingModel output is rendered/executed without sanitisationTreat model output as untrusted; encode and validate
Supply chainA compromised model, plugin, or dataset enters your stackVet providers and plugins; pin versions; review dependencies

The remaining items (training data poisoning, model DoS, insecure plugin design, model theft) matter more for teams training or hosting their own models.

A sensible rollout order

  1. Prompt injection and sensitive disclosure first — these affect almost every LLM feature.
  2. Excessive agency next, if you are building agents.
  3. Overreliance for anything user-facing with consequences (health, finance, legal).

What this is not

OWASP is a checklist, not a compliance certificate. It does not replace your data-protection, security, or accessibility obligations. It is, however, a far better starting point than a generic web-app checklist.

OWASP LLM Top 10: rollout order 1. Prompt injection + data disclosure 2. Excessive agency (agents) 3. Overreliance (user-facing) Checklist, not a certificate start here, then go deeper

VERDICT

Start with prompt injection and data disclosure, then agency and overreliance. OWASP's list is a free, practical starting point, not a substitute for your own controls.

Sources