Found during a routine audit of main at 888b83c926
Description
Found during a routine audit of main at 888b83c926. Reported privately per SECURITY.md. Reproduced, output verbatim below.
The weakness
The MCP permission model has a distinct allowSchemaChange flag, which an operator sets precisely to narrow an agent below what its database user could do. ExecuteCommandTool gates on it through the engine's operation-type analysis. The two upsert tools do not.
upsert_entity and upsert_relationship build a Cypher MERGE and run it through MCPToolUtils.executeParameterizedWrite (mcp/src/main/java/com/arcadedb/mcp/tools/MCPToolUtils.java:309-325), which analyzes the generated statement and gates it the same way. A MERGE naming a type that does not exist analyzes as {CREATE, UPDATE} and not as a schema change, but executing it creates the type.
So a caller holding insert and update, and explicitly denied schema change, creates arbitrary new types.
The tool descriptions advertise the auto-creation ("Created automatically if it does not exist", UpsertEntityTool.java:52, UpsertRelationshipTool.java:53), so the behaviour is intended. What is not intended is that it is not gated.
Reproduction
allowSchemaChange=false, allowInsert and allowUpdate true, root as the caller so that no engine-side grant is in the way and the only gate under test is the MCP one:
PROBE-MCP type exists before = false
PROBE-MCP execute_command CREATE VERTEX TYPE -> SecurityException: Schema change operations are not allowed by MCP configuration
PROBE-MCP upsert_entity -> {"records":[{"@rid":"#28:0","@type":"AuditR25NewType","@cat":"v","id":"x1"}],"count":1}
PROBE-MCP type exists after = trueThe same server, the same flag, the same schema change: refused through one tool and performed through another.
Refutation attempted
Does the ArcadeDB-side ACL still stop a non-root caller? It stops a caller without the engine grant. It does not restore allowSchemaChange, which is a separate MCP-level restriction with its own documented purpose.
Is the type created only because root was used? No. The gate under test runs before the engine ACL and is identical for every caller. Root only removes the second, unrelated gate from the picture.
Could analyze() be expected to report SCHEMA here? It cannot know: whether the MERGE creates a type depends on the schema at execution time, not on the statement text. That is why the fix belongs in the tool and not in the analyzer.
Fix, and the guarantee
In both upsert tools, check whether each named type, and for upsert_relationship the edge type, already exists before building the statement, and require allowSchemaChange when one does not. Likewise for a property the SET clause would add, if the MCP model intends to cover properties.
The guarantee we would ask for: a test that enumerates every tool able to reach the engine with a write, taken from MCPDispatcher's own registry rather than a hand-written list, and asserts that with allowSchemaChange=false no tool leaves the schema different from how it found it, comparing the full schema before and after. A per-tool assertion on a single statement shape is what let these two through.
Scope
Executed for upsert_entity creating a vertex type, output above. upsert_relationship was read and not run: it builds two node MERGEs and an edge MERGE through the same appendNodeMerge and executeParameterizedWrite path (UpsertRelationshipTool.java:88-100), so the same gap is expected for its vertex and edge types, but we did not drive it. Whether a property added by the SET clause counts as a schema change under this model was also not measured.
Dedup
Nothing matching in the last 1000 issue titles. GHSA-6x73-v3rc-f57c (principal binding) and GHSA-pff6-hp53-pj54 (server-admin tools) are the two MCP authorization advisories already fixed. Both are about who the caller is; this is about which operation the configuration permits.
Related advisories
This issue is tracked together with GHSA-r946-225c-5x88, which carries the CVE identifier for the group.