Auth0 AI FGA retrievers let documents denied by Auth0 FGA into the RAG context (CVE-2026-108167, CVE-2026-107372)
The FGA retrievers in Auth0's AI SDKs, which filter retrieved documents by Auth0 FGA permissions before they reach the model, could keep documents the user was denied. When several documents in one retrieval map to the same FGA object but differ by user, relation or condition context, one authorization result was reused for all of them. Upgrade auth0-ai-langchain to 1.0.2, auth0-ai-llamaindex to 1.0.3, and @auth0/ai to 6.0.3.
What it is
CVE-2026-108167 (Python, auth0-ai-langchain and auth0-ai-llamaindex): the retriever does not distinguish authorization checks under different users, relations or conditions, so documents denied by FGA are included in the model context and used in the answer. It applies when one retrieval returns several documents that map to the same FGA object with different users, relations or authorization contexts.
CVE-2026-107372 (JavaScript, @auth0/ai): the same flaw for conditional checks. It applies when the retriever uses FGA conditions and one retrieval returns several documents with the same object, user and relation but different condition context.
Both are CWE-863 Incorrect Authorization with confidentiality impact only (CVSS 6.5, attacker needs a low-privilege account). Okta published both advisories on 10 Nov 2026.
How to check
- Python: pip show auth0-ai-langchain auth0-ai-llamaindex and compare with 1.0.2 / 1.0.3.
- JavaScript: npm ls @auth0/ai and compare with 6.0.3.
- Check whether you use the FGA retriever and whether several chunks or documents share one FGA object (for example chunks of the same file) with different users, relations or conditions.
How to fix
- Upgrade to auth0-ai-langchain 1.0.2+, auth0-ai-llamaindex 1.0.3+, and @auth0/ai 6.0.3+.
- If any users could have seen documents they were denied, review what the assistant answered from restricted sources.
pip install -U 'auth0-ai-langchain>=1.0.2' 'auth0-ai-llamaindex>=1.0.3'
npm install @auth0/ai@^6.0.3 Sources
- criticalCVE-2026-75128: Langchain-Chatchat lets anyone run shell commands through /tools/call
- criticalCVE-2024-46946: LangChain Experimental Eval Injection vulnerability
- criticalCVE-2024-27444: LangChain Experimental vulnerable to arbitrary code execution
- criticalCVE-2023-32785: Langchain SQL Injection vulnerability