Vulnerability  ·  2026-09-19

ArcadeDB AI chat endpoint — query_database tool bypasses per-type/per-bucket ACLs

VulnerabilityMedium impactGlobalCVE-2026-93595
ArcadeDB's LLM-facing chat/query tool fails to propagate the authenticated user's identity into the database access-control context, so ACL restrictions configured for that user are silently bypassed when queries run through the AI chat interface instead of the standard API.
This is a textbook example of an AI tool integration inadvertently creating a privilege-escalation side channel: a user who is properly restricted via ACLs on the normal database API can use the AI assistant to read data they should not have access to, undermining the database's core access-control model.
The query_database tool exposed through ArcadeDB's AI chat endpoints executes queries without binding the authenticated principal to the DatabaseContext, causing per-type and per-bucket ACL checks to silently no-op, letting users query and read data they are not authorized to access via the AI assistant interface.
ArcadeDB before 26.9.1
Upgrade to ArcadeDB 26.9.1 or later per GHSA-chrr-vr3p-crcc.
GitHub Security Advisory GHSA-chrr-vr3p-crccNVD CVE-2026-93595
See this in the live feed Explore related AI security and governance findings — updated every morning.
Open the feed →