CVE-2026-93328: Google GenAI MCP Toolbox Credential Leak
A connection-string validation and metadata reflection flaw in MCP Toolbox for Databases allows crafted Model Context Protocol tool invocations to exfiltrate upstream database credentials and cloud service account tokens to untrusted model contexts.
Executive Summary & Technical Impact
A connection-string validation and metadata reflection flaw in MCP Toolbox for Databases allows crafted Model Context Protocol tool invocations to exfiltrate upstream database credentials and cloud service account tokens to untrusted model contexts.
Tracked under CVE-2026-93328 (CWE-200 (Exposure of Sensitive Information) / SSRF), this vulnerability affects MCP Toolbox (genai-toolbox) 1.2.0 – 1.4.0 (< 0.8.4 / < 1.5.0) and carries a severity score of 9.3 (CRITICAL). Engineering and security operations teams running MCP Toolbox for Databases (genai-toolbox) should prioritize upgrading to MCP Toolbox v1.5.0 / genai-toolbox v0.8.4.
Remediation & Patch Guidance
- Upgrade genai-toolbox / MCP Toolbox for Databases to version 1.5.0 (or v0.8.4+) immediately.
- Bind MCP server listeners exclusively to localhost or authenticated mTLS internal service meshes.
- Rotate database credentials and IAM service account tokens previously passed to vulnerable MCP Toolbox instances.
Frequently Asked Questions (CVE-2026-93328)
What versions are affected by CVE-2026-93328?
CVE-2026-93328 affects MCP Toolbox (genai-toolbox) 1.2.0 – 1.4.0 (< 0.8.4 / < 1.5.0) in MCP Toolbox for Databases (genai-toolbox) (Google GenAI MCP Toolbox Credential Leak).
How do I patch or remediate CVE-2026-93328?
Upgrade affected deployments to MCP Toolbox v1.5.0 / genai-toolbox v0.8.4 and verify the remediation steps listed above.
What is the CVSS score of CVE-2026-93328?
CVE-2026-93328 is rated CRITICAL with a CVSS score of 9.3 (vector: CVSS:3.1/AV:N/AC:L/PR:N/UI:N/S:C/C:H/I:L/A:N).
Critical CVEs Hit AI Agent and Cloud Infrastructure: OpenHands SDK RCE, Google MCP Toolbox File Exfiltration, and Consul-k8s ACL Bypass
Three critical advisories published on October 26, 2026 share one lesson: default-open settings in agent sandboxes, MCP servers, and service meshes are an attack surface.
Read full engineering post-mortem (6 min read) →