Skip to content
OWASP LLM Top 10 guide

LLM09:2025

Misinformation

Misinformation occurs when an AI system produces false, unsupported, outdated, or misleading output that users or downstream systems treat as reliable.

Step 01

Input

Step 02

Model

Step 03

Tool / Data

Step 04

Impact

What it is

The model produces a false or misleading answer that appears credible. Missing context, fabricated details or biased source material can contribute; a citation is useful only if it supports the claim being made.

Why it matters

A customer may act on an incorrect policy answer, an engineer may use an invented package, or a reviewer may accept a false claim in a report. Consequences depend on the decision made from the answer.

Failure path

One way this failure can develop.

Use this example to locate the input, permission or application check that would interrupt the sequence in your system.

Path 01

Ask questions where the model lacks reliable context or current facts.

Path 02

Exploit weak grounding, missing citations, ambiguous prompts, or overconfident response style.

Path 03

Push users or systems to act on unsupported output.

Defenses

Controls worth checking.

Check where each control runs, who owns it and how you will test it again after a prompt, model or workflow change.

Control 01

Ground high-impact answers

Require citations, source snippets, freshness indicators, and answer abstention for workflows where unsupported output creates risk.

Control 02

Constrain claims

Define which claims require a source check or qualified review before use. Test abstention on missing or conflicting information; a model's confident wording is not a reliability measure.

Control 03

Design for uncertainty

Make the product show what is known, what is inferred, and what needs human review.

Signals and control gaps

  • Answers with no source for factual claims.
  • Overconfident output when retrieval returned weak or empty context.
  • Users copying AI output into high-impact workflows without review.

Questions for your team

  • Which answers require source grounding?
  • Does the model clearly abstain when context is missing?
  • Where could a wrong answer create a customer, security, or legal issue?