AI Security Atlas › All entries › LLM Application & Output Security
Grounding & Citation Verification
Answers are built from retrieved sources, each claim carries a citation, and citations are checked before the answer is shown.
Open in the interactive atlas →Impact
Turns silent hallucinations into visible abstentions or checkable claims.
Why this severity
Guardrails take the highest severity of the risks they mitigate: Ungrounded, Unverified Answers (High), Hallucinated Answers Acted On (High).
Method: OWASP Risk Rating Methodology
Related
Mitigates
- Ungrounded, Unverified Answers Vulnerability · High severity
- Hallucinated Answers Acted On Threat · High severity
Needed by
- Uncertainty Labeling & High-Stakes Review Guardrail · Fixes high risks
Remediation
- Retrieve from a curated, versioned knowledge base.
- Check that every citation resolves and supports its claim.
- Prefer 'I don't know' over an unsupported answer.
How to verify
- Accuracy and citation-validity metrics gate each release.
- Fabricated-citation fixtures are rejected by the checker.
Framework mapping
OWASP Top 10 for LLM Applications 2025
- LLM09:2025 Misinformation
MITRE ATLAS
NIST AI RMF 1.0
- MEASURE 2.5 AI system is demonstrated to be valid and reliable