Vector and embedding weaknesses include index poisoning, unauthorized retrieval, and similarity abuse that surfaces sensitive data. Attackers manipulate retrieval context to bias responses or exfiltrate records through crafted queries. Weak access boundaries and poor corpus hygiene significantly increase impact.
Apply strict access controls and tenant isolation to vector infrastructure. Validate retrieval requests, harden indexing pipelines, and monitor retrieval anomalies for abuse indicators. Periodically test corpus integrity and relevance controls to ensure retrieval remains trustworthy and policy-compliant.
The same feature written the risky way, then with the fix applied.
How this risk has ranked in each OWASP release.
2025Latest
#08llm08-2025 Vector and Embedding Weaknesses
