MCP database server back-end validation flaws expose SQL and metadata paths
Akamai research describes three database MCP server failures: SQL injection in Apache Doris MCP, SQL-capable unauthenticated paths in Apache Pinot MCP, and unauthenticated metadata exposure in Alibaba RDS MCP.
Date
May 12, 2026
First Seen
May 12, 2026
Last Reviewed
May 17, 2026
Publisher
Akamai
Source Type
article
Related reading
OpenClaw Security GuideA practical baseline for local binding, scoped credentials, sandboxing, runtime checks, and Armorer Guard.
Securing OpenClaw with Armorer GuardHow Armorer wraps OpenClaw with managed setup, Docker hardening, health checks, approvals, and Guard-backed scanning.
MCP Database Server Back-End Validation Flaws Expose SQL and Metadata Paths
Summary
Akamai research describes three database MCP server failures: SQL injection in Apache Doris MCP, SQL-capable unauthenticated paths in Apache Pinot MCP, and unauthenticated metadata exposure in Alibaba RDS MCP.
Why It Matters
Agent operators often treat MCP servers as simple adapters. In practice, a database MCP server can become a privileged gateway into production data, schema metadata, and cloud-managed database APIs. If it binds to a reachable network interface, accepts unauthenticated clients, or passes partially validated parameters into SQL or retrieval code, the model-facing tool boundary becomes a data-plane security boundary.
For OpenClaw and Armorer-style local agent stacks, the durable lesson is that connector inventory and runtime policy need to include the back-end side of a tool, not only the prompt/tool-description side. A connector that appears safe in the agent UI can still hold broad database credentials, expose HTTP/SSE listeners, or perform unsafe query construction internally.
Attack Path
- An MCP database server is deployed with SQL-capable tools, schema-retrieval tools, or RAG-backed metadata retrieval.
- The server is reachable by an untrusted client, compromised agent session, or lower-trust network segment.
- Authentication is missing, optional, or only applied to some paths; alternatively, one tool parameter bypasses validation.
- A caller invokes an exposed MCP tool with crafted input or broad query intent.
- The MCP server relays the request to the back-end database or vector index using its own privileges, enabling SQL injection, unauthorized queries, metadata disclosure, or broader database impact depending on the integration.
Affected Surface
- Apache Doris MCP Server v0.6 and earlier, according to Akamai, for
CVE-2025-66335SQL injection through thedb_nameparameter. - Apache Pinot MCP v1.1.0 and earlier, according to Akamai, where reachable HTTP transport and weak query validation could expose SQL-capable tools; StarTree added OAuth as an option in later work.
- Alibaba RDS MCP, according to Akamai, where the FastMCP-based RAG component exposed table-structure retrieval without authentication and Alibaba declined to patch as of the reported disclosure timeline.
- Any agent environment where MCP database connectors are reachable beyond a tightly controlled local trust boundary or where tool-server credentials exceed least privilege.
Armorer Relevance
Armorer could help reduce this class of risk by giving operators a local control plane for MCP and tool-server inventory: which connectors are enabled, where they bind, what credentials they receive, and which back-end systems they can reach. Docker isolation can limit host filesystem and network paths for connector processes. Runtime monitoring can flag unexpected listeners, outbound database connections, tool-call volume spikes, and suspicious child processes. Credential handling can keep database tokens scoped and isolated from the broader agent environment. Health checks could warn when MCP servers bind to 0.0.0.0, expose HTTP/SSE without authentication, or run with overly broad database privileges.
These controls would not prove prevention of the Akamai-reported vulnerabilities, but they would make unauthenticated transports, unsafe exposure, and oversized connector privileges harder to miss and easier to contain.
Evidence
- Akamai published the research on 2026-05-12 and described the common pattern as missing or faulty validation between MCP servers and their back ends.
- Akamai identifies Apache Doris MCP v0.6 and earlier as affected by
CVE-2025-66335, with a patch issued by Apache on 2025-12-30. - Akamai identifies Pinot MCP v1.1.0 and earlier as affected by unauthenticated HTTP transport and weak SELECT validation in reachable deployments, with OAuth added as a later authentication option.
- Akamai identifies Alibaba RDS MCP as exposing unauthenticated table-structure retrieval from a vector index and says Alibaba marked the issue not applicable for a fix.
Mitigations
- Patch affected MCP database servers and verify versions against primary advisories or project release notes.
- Require authenticated transport for HTTP, SSE, WebSocket, and other remotely reachable MCP endpoints.
- Bind MCP servers to localhost or trusted networks unless remote access is explicitly required and protected.
- Validate every tool parameter server-side before query construction; avoid partial query validation that only inspects the first parsed statement or top-level intent.
- Give MCP connectors least-privilege database credentials and separate credentials per connector where practical.
- Monitor MCP server listeners, tool calls, database query patterns, and outbound network activity.
- Treat newly added or updated MCP connectors as privileged code changes requiring review.
Open Questions
- Which downstream packages or managed deployments embedded the affected MCP database servers.
- Whether the Pinot query-validation issue receives a full fix beyond optional OAuth transport authentication.
- Whether CERT/CC or Alibaba publish additional guidance for the RDS MCP metadata exposure.
- Which Armorer health checks should become default for database-backed MCP servers.