CVE-2026-63129: mssql-mcp-core: SQL injection in mssql-mcp write tools via unparameterized table names and WHERE clauses
The mssql-mcp-core MCP packages for Microsoft SQL Server built T-SQL by pasting table names, column names and WHERE clauses into queries. A malicious MCP client, or an LLM agent steered by prompt injection, can run arbitrary SQL, including stacked statements. Upgrade @connorbritain/mssql-mcp-core to 0.6.0, reader and writer to 0.5.0, and server to 0.6.0.
Published Nov 19, 2026 · Analysis Nov 19, 2026
What an attacker can do
Several tools passed caller-supplied identifiers and predicates straight into the node-mssql Request.query() batch API. Only data values were parameterized. update_data, delete_data and execute_transaction used a raw tableName and a free-form whereClause. insert_data used raw column names. create_table, create_index and drop_table had weak identifier checks. list_databases interpolated the stateFilter argument, and that tool belongs to the read-only tier.
The only check on whereClause was that it was not empty. The row-count and preview queries run before the confirmUpdate or confirmDelete step. A hostile whereClause or tableName therefore ran on the first call and skipped the preview, the confirmation and the maxRows limit.
Request.query() accepts multi-statement batches. An attacker who controls tool arguments can run stacked DDL and DML, read or change any data the login can reach, and drop tables. They can also get around the read_data allow and deny validation. If the SQL login is privileged and xp_cmdshell is enabled, they can run operating system commands. In practice the attacker is rarely the MCP client itself. More often a web page, ticket or document that the agent reads carries instructions, and the agent turns them into tool arguments.
Who is affected
These npm packages are affected: @connorbritain/mssql-mcp-core up to 0.5.0, @connorbritain/mssql-mcp-reader and @connorbritain/mssql-mcp-writer up to 0.4.2, and @connorbritain/mssql-mcp-server up to 0.5.2. The reader package is affected too, even with READONLY=true, because of list_databases. Check what you have installed:
npm ls @connorbritain/mssql-mcp-core @connorbritain/mssql-mcp-reader @connorbritain/mssql-mcp-writer @connorbritain/mssql-mcp-server@connorbritain/mssql-mcp-core
- Ecosystem
- npm
- Affected
- <= 0.5.0
- Fixed in
- 0.6.0
- Check
npm ls @connorbritain/mssql-mcp-core
@connorbritain/mssql-mcp-reader
- Ecosystem
- npm
- Affected
- <= 0.4.2
- Fixed in
- 0.5.0
- Check
npm ls @connorbritain/mssql-mcp-reader
@connorbritain/mssql-mcp-writer
- Ecosystem
- npm
- Affected
- <= 0.4.2
- Fixed in
- 0.5.0
- Check
npm ls @connorbritain/mssql-mcp-writer
@connorbritain/mssql-mcp-server
- Ecosystem
- npm
- Affected
- <= 0.5.2
- Fixed in
- 0.6.0
- Check
npm ls @connorbritain/mssql-mcp-server
How to fix it
Upgrade to the patched versions: @connorbritain/mssql-mcp-core 0.6.0, @connorbritain/mssql-mcp-reader 0.5.0, @connorbritain/mssql-mcp-writer 0.5.0 and @connorbritain/mssql-mcp-server 0.6.0. The consumer packages depend on core ^0.6.0.
npm install @connorbritain/[email protected]The fix validates all identifiers against a strict allowlist and bracket-quotes them. It also removes the free-form whereClause parameter, which is a breaking change. Callers now send structured filters of the form { column, operator, value }, with an operator allowlist and bound values. stateFilter is checked against a fixed list. The advisory says upgrading is the only complete fix. As defense in depth, run the server under a least-privilege SQL login with xp_cmdshell disabled, and limit which MCP clients can call the tools.
How to tell if you were hit
Review SQL Server logs and any query tracing for batches from the MCP server's login that contain stacked statements, DROP, xp_cmdshell, or odd text in WHERE clauses and table names. Look for list_databases calls whose stateFilter is not one of the normal state values. Check for unexpected tables, indexes or schema changes. Also check whether xp_cmdshell was enabled on the server. Beam records the MCP tool calls an agent makes, so you can search them for tool arguments that carry SQL fragments.
Original advisory
From the GitHub Advisory Database (CC BY 4.0).
Several tools in mssql-mcp-core built T-SQL by string-interpolating caller-supplied identifiers and predicates directly into the statement passed to the node-mssql Request.query() batch API. Only data values were parameterized.
Affected tools and inputs:
update_data,delete_data:tableNameand the free-formwhereClausewere interpolated raw.insert_data:tableNameand column names (object keys) were interpolated raw.execute_transaction: the same rawtableName/whereClauseinterpolation per operation.create_table,create_index,drop_table: table/column/index/schema identifiers were interpolated with no or insufficient validation.list_databases(reader tier): thestateFilterargument was interpolated raw into the query, with only a non-enforced inputSchema enum guarding it.
The only check on whereClause was that it was non-empty. Because the row-count and preview queries run before the confirmUpdate/confirmDelete step, a malicious whereClause or tableName executed on the first call, bypassing the preview/confirmation and maxRows safety controls. Request.query() submits a multi-statement T-SQL batch, so a crafted input enabled stacked statements (arbitrary DDL/DML), sub-selects, and (with a privileged login and xp_cmdshell enabled) OS command execution. Because list_databases is a read-only tool, the stateFilter sink makes the reader package injectable as well.
Impact
An actor who can supply tool arguments (a malicious or compromised MCP client, or an LLM agent steered via indirect prompt injection) could read, modify, or drop arbitrary data, bypass the read_data allow/deny validation and the preview/confirmation/maxRows safeguards, and conditionally achieve OS command execution.
Patches
Fixed in @connorbritain/mssql-mcp-core 0.6.0:
- All identifiers are validated against a strict allowlist and bracket-quoted via a shared helper.
- The free-form
whereClausestring parameter is removed (breaking change) and replaced with structured, parameterized filters ({ column, operator, value }) using an operator allowlist. Values are always bound, never interpolated. list_databasesvalidatesstateFilteragainst a fixed allowlist before use.
Upgrade the consumer packages: mssql-mcp-reader 0.5.0, mssql-mcp-writer 0.5.0, mssql-mcp-server 0.6.0 (all depend on core ^0.6.0).
Workarounds
Upgrading is the only complete fix; the list_databases (stateFilter) sink is reachable even under READONLY=true and in the reader package. As defense in depth, run the server under a least-privilege SQL login with xp_cmdshell disabled, and restrict which MCP clients can invoke the tools.
Credited to novice-22, m10x.
Questions
- Does the read-only mode protect me from this SQL injection?
No. The
list_databasestool is in the reader tier, and itsstateFilterargument was interpolated raw. It is reachable underREADONLY=trueand in the reader package. Upgrade to fix it.- Can this lead to operating system command execution?
Only conditionally. The advisory says it needs a privileged SQL login with
xp_cmdshellenabled. Without that, the attacker is limited to what the SQL login can do in the database, which can still mean reading, changing or dropping data.- Will upgrading to 0.6.0 break my existing prompts or scripts?
It can. The free-form
whereClauseparameter is removed. Tools now take structured filters withcolumn,operatorandvaluefields. Update any client that sent raw WHERE strings.
References
Read for this analysis:
- https://nvd.nist.gov/vuln/detail/CVE-2026-63129
- https://github.com/ConnorBritain/mssql-mcp-core/pull/1
- https://github.com/ConnorBritain/mssql-mcp-core/commit/227e2efc4e3311c477b7ae509c3ae15437a4f815
- https://github.com/ConnorBritain/mssql-mcp-core/commit/a9727f83768598e4fdbddfbc717ccfebb84ae7af
- https://github.com/advisories/GHSA-2m9m-6cr5-9x25
How this page is made: the data panel comes from the GitHub Advisory Database, CISA’s KEV catalog, the NVD and FIRST EPSS, and updates twice a day. The analysis is written by an AI model from the sources listed above and checked by automated rules before it is published. Check versions against the vendor's advisory before you act. Methodology
More Other MCP servers and tools vulnerabilities
All Other MCP servers and tools entriesCVE-2026-77312
Flyto2 Core: Arbitrary file write via unguarded data.*/file.* modules (incomplete fix of CVE-2026-67429)
critical10.0Nov 20, 2026CVE-2026-77318
Jupyter MCP Server: Unauthenticated management routes can hijack the Jupyter backend in standalone HTTP mode
high7.3Nov 20, 2026CVE-2026-77359
Jupyter MCP Server: Browser DNS rebinding can hijack the Jupyter backend through management routes
critical9.3Nov 20, 2026CVE-2026-130007
k8s-mcp-server: Missing authentication on default SSE/HTTP transport allows unauthenticated Kubernetes tool execution
high7.9Nov 19, 2026CVE-2026-130008
sqlite-mcp: Read-only mode bypass via statement chaining and SQL injection in describe_table
high8.2Nov 19, 2026CVE-2026-62964
Gortex: Arbitrary file write / repo-confinement escape in `write_file` / `edit_file` / `move_inline` MCP tools → code execution
high7.8Nov 19, 2026
