Ungrounded, Unverified Answers

Responses are generated from model memory alone, with no retrieval, citation, or check that cited sources exist and support the claim.

Open in the interactive atlas →

Impact

Fabricated facts and citations reach users with nothing to catch them.

Why this severity

Impact Medium × Likelihood High = High

Impact medium: errors are harmful only when relied on. Likelihood high: ungrounded answers caused the Air Canada liability ruling.

Impact Medium: Limited or recoverable harm: cost, availability, reputational damage, or exposure confined to one session or user.
Likelihood High: At least one documented real-world incident, and the attack needs no access beyond what the system exposes to ordinary users or content (publishing to a public registry counts as ordinary).

Method: OWASP Risk Rating Methodology

Evidence

Incident: Moffatt v. Air Canada, 2024 BCCRT 149: airline held liable for its chatbot's wrong bereavement-fare answer (BC Civil Resolution Tribunal decision (PDF), 2024)

Exploited by

Hallucinated Answers Acted On Threat · High severity

Mitigated by

Grounding & Citation Verification Guardrail · Fixes high risks

Remediation

Answer from retrieved documents and require a citation for each factual claim.
Verify that cited documents exist and contain the cited passage.
Let the model say it does not know; reward abstention in evals.

How to verify

Citation checker rejects answers whose quotes are absent from the source.
Measure unsupported-claim rate on a held-out question set.

Pattern

const ok = claims.every(c => sources[c.ref]?.text.includes(c.quote));
if (!ok) return abstain('unsupported claim');

Framework mapping

OWASP Top 10 for LLM Applications 2025

LLM09:2025 Misinformation

MITRE ATLAS

AML.T0062 Discover LLM Hallucinations

NIST AI RMF 1.0

MEASURE 2.5 AI system is demonstrated to be valid and reliable