Vulnerability  ·  2026-11-20

mssql-mcp-core SQL injection in write/DDL/transaction tools (conditional xp_cmdshell)

VulnerabilityHigh impactGlobalCVE-2026-63129
The shared library that powers MSSQL MCP servers interpolated caller-controlled identifiers and the free-form whereClause into T-SQL batches, parameterizing only data values. The fix introduces sqlIdentifier.ts (allowlist + bracket-quoting) and whereFilter.ts (operator allowlist, always-bound values) and applies them across update_data, delete_data, insert_data, execute_transaction, create_table, create_index and drop_table. Count/preview queries also ran before the confirmation gate.
MSSQL MCP servers let LLM agents issue database operations. Because injection fires on unconfirmed first calls and is reachable by prompt-injected agents, a single malicious action can escalate to arbitrary SQL execution (and via xp_cmdshell, code execution) against the database the agent is trusted to query.
Tool arguments (table/column/index identifiers and the free-form whereClause) are string-interpolated into T-SQL passed to Request.query() while only data values are parameterized, so a caller — or an LLM steered by prompt injection — can inject stacked statements, DDL, or conditional xp_cmdshell on reachable SQL Server instances.
mssql-mcp-core below 0.6.0 (mssql-mcp server below 0.6.0, reader/writer below 0.5.0)
Upgrade to mssql-mcp-core/server 0.6.0 (reader/writer 0.5.0), which add an identifier allowlist/bracket-quoting and structured parameterized filters; note the breaking change to a structured filters array.
mssql-mcp-core SQL injection fix commitNVD — CVE-2026-63129
See this in the live feed Explore related AI security and governance findings — updated every morning.
Open the feed →