
Auth0 AI SDK patches are out: @auth0/ai 6.0.3 for JS, auth0-ai-langchain 1.0.2 and auth0-ai-llamaindex 1.0.3 for Python. If Auth0 FGA decides which documents your RAG app can show a user, take them today 🔐 The bug (CVE-2026-107372 JS, CVE-2026-108167 Python) The FGA retriever checks every document before it reaches the model. But when one retrieval pulled several documents that point at the same FGA object under different users, relations or conditions, their checks got treated as one. A document FGA had denied could still land in the context and end up in the answer The JS fix says it all: the check key now includes the context and contextual tuples Lesson if you cache authorization results: the cache key has to carry every input to the decision, not just the object it points at Upgrade npm i @auth0/ai@latest pip install -U auth0-ai-langchain auth0-ai-llamaindex https://trust.okta.com/security-advisories/improper-access-control-in-auth0-ai-python-document-retrieval-cve-2026-108167
Post summary
The post announces patched Auth0 AI SDK versions for CVE-2026-107372 and CVE-2026-108167 and urges affected users to upgrade. It also briefly explains the authorization-cache flaw that could allow denied documents into RAG context.
