Disclaimer: This content was generated using AI. While we strive for accuracy, please verify critical information with official sources.
Sources: National Vulnerability Database, CISA KEV, Exploit-DB
CVE-2026-115324 Overview
CVE-2026-115324 is a SQL injection vulnerability in the mem0 memory layer for AI agents, affecting versions 0.1.117 through 1.0.7. The flaw resides in mem0/vector_stores/databricks.py, where insert() and update() operations build SQL statements through unescaped string interpolation. Attackers can inject arbitrary SQL by supplying crafted memory text, metadata fields, or message names that reach the Databricks SQL warehouse. Because mem0 stores content supplied by end users and upstream large language models (LLMs), untrusted input readily flows into the vulnerable code paths. The issue is fixed in mem0 1.0.8 (release notes).
Critical Impact
Attackers can execute arbitrary SQL against the configured Databricks SQL warehouse, enabling data theft, modification, or destruction of memory stores backing AI applications.
Affected Products
- mem0 versions 0.1.117 through 1.0.7 using the Databricks vector store backend
- Applications relying on mem0/vector_stores/databricks.py for memory persistence
- Downstream AI agent frameworks that embed vulnerable mem0 releases
Discovery Timeline
- 2026-10-22 - CVE-2026-115324 published to NVD
- 2026-10-22 - Last updated in NVD database
Technical Details for CVE-2026-115324
Vulnerability Analysis
mem0 provides a long-term memory layer for AI agents, persisting conversation context into pluggable vector stores. The Databricks backend constructs SQL INSERT and UPDATE statements by interpolating Python values directly into query strings. User-controlled fields including memory text, metadata values, and message names are embedded without escaping or parameter binding. An attacker who can influence any content written into mem0 can terminate the intended statement and append arbitrary SQL executed on the Databricks SQL warehouse.
The attack surface is broad because mem0 is designed to ingest unstructured input from users and LLM outputs. In agent architectures, model-generated memory writes effectively act as a confused deputy, carrying prompt-injected SQL payloads into the data layer.
Root Cause
The root cause is improper neutralization of special elements used in an SQL command [CWE-89]. The Databricks integration relies on Python string interpolation rather than parameterized queries or the Databricks SQL connector's parameter binding features. Any single quote, semicolon, or SQL keyword inside input fields alters query structure.
Attack Vector
Exploitation requires the attacker to deliver crafted text into any workflow that calls mem0's add() or update() APIs backed by Databricks. This can occur through direct user chat input, document ingestion, tool outputs, or indirect prompt injection via retrieved content. See the VulnCheck SQL Injection Advisory and the upstream patch commit for technical details on the vulnerable code paths.
Detection Methods for CVE-2026-115324
Indicators of Compromise
- Databricks SQL warehouse query history containing memory-store statements with unexpected UNION, ;, or comment sequences (--, /*).
- mem0 application logs showing insert() or update() calls with payloads containing unescaped single quotes or SQL keywords in memory text or metadata.
- Unexpected schema changes, new tables, or anomalous row counts in Databricks tables used by mem0.
- Service account activity on the Databricks warehouse outside of normal mem0 operation windows.
Detection Strategies
- Enable Databricks SQL warehouse query logging and alert on syntactically anomalous statements originating from the mem0 service principal.
- Perform software composition analysis across Python environments to identify mem0 releases prior to 1.0.8.
- Instrument the mem0 client to log raw memory payloads and flag entries containing SQL metacharacters for review.
Monitoring Recommendations
- Forward Databricks audit logs and mem0 application logs into a centralized analytics platform for correlation.
- Baseline the normal set of SQL statement templates issued by mem0 and alert on deviations.
- Monitor outbound data volumes from the Databricks warehouse associated with the mem0 service account.
How to Mitigate CVE-2026-115324
Immediate Actions Required
- Upgrade mem0 to version 1.0.8 or later across all environments using the Databricks vector store.
- Rotate Databricks personal access tokens and service principal secrets used by mem0 deployments.
- Review Databricks query history for the exposure window and investigate any anomalous statements.
- Audit mem0-backed tables for unauthorized rows, dropped objects, or schema modifications.
Patch Information
The fix is included in mem0 1.0.8, published in the GitHub Release v1.0.8. The remediation is tracked in the upstream commit and pull request #4558, with background in issue #4073. The patch replaces string interpolation with safe parameter handling in the Databricks vector store module.
Workarounds
- Switch to a different mem0 vector store backend until the Databricks integration is upgraded.
- Restrict the Databricks service principal used by mem0 to least-privilege scope on dedicated schemas, preventing cross-table access.
- Apply input validation at the application layer to reject memory content containing SQL metacharacters where feasible.
- Isolate the Databricks SQL warehouse behind network controls that limit access to the mem0 application tier only.
Disclaimer: This content was generated using AI. While we strive for accuracy, please verify critical information with official sources.
Sources: National Vulnerability Database, CISA KEV, Exploit-DB

