Skip to content

Security FAQ

Last updated

View as Markdown

These controls limit what DBGorilla, and engineers using DBGorilla, can do to a production database. They cover the collector, the MCP integration, natural-language queries and the CI review bot on GitHub and GitLab. The current status of each control is in the Control status table.

  • DBGorilla has no direct network path to your database. Everything goes through the collector, which runs in your network, connects with a read-only role you create, and makes outbound connections only.
  • The agents interact with your database only through a fixed set of tools. None of those tools can write to your database. One, run_experiment, runs code with full rights on a temporary clone that holds your schema and optimizer statistics but no rows.
  • You can turn off command execution entirely from the collector config, in your infrastructure, and DBGorilla cannot turn it back on.
  • Anything DBGorilla proposes is a recommendation. A person still reviews and applies every change.

The collector connects with a dedicated read-only database role that you create and control. The grants are pg_monitor, plus USAGE and SELECT on the schemas you want captured, which is the same class of access a database monitoring agent already has in most environments. The collector cannot escalate beyond the privileges of that role. The exact grants are in Prepare PostgreSQL.

What does the collector read, and what load does it add?

Section titled “What does the collector read, and what load does it add?”

Host, network, storage and I/O metrics, query statistics from pg_stat_statements, topology and schema. The additional load on the instance is negligible; in most environments it reads the same statistics that existing monitoring already reads.

Can the agent write to my database through the collector?

Section titled “Can the agent write to my database through the collector?”

No. The agent only interacts with your database through tools, and none of the tools it has can write. There is no tool for arbitrary SQL, no tool for DDL, and no tool that modifies or removes data. The collector is the single choke point for everything DBGorilla reads or runs, and it is entirely under your control.

Can command execution be disabled entirely?

Section titled “Can command execution be disabled entirely?”

Yes. The [commands] block in collector.toml is optional and lives in your infrastructure. With commands off, the collector will not execute any command on your behalf: the agents are limited to the metrics and metadata the collector reports. When commands are on, they permit three read-oriented operations:

  • explain: plan-only EXPLAIN. The query is never executed.
  • execute_query: a single SELECT in a read-only transaction that is always rolled back, capped at 1,000 rows with a 30-second statement timeout.
  • collect_statistics: captures optimizer statistics for named tables, and no table rows. On PostgreSQL the statistics include each column’s most common values and histogram boundaries, which are values from your data, and code run in an experiment can read them.

enabled is the default for components that do not decide for themselves. A [[component]] carrying its own commands list uses that list and ignores the switch, in both directions, so to turn commands off everywhere set enabled = false and leave commands unset on every component. Commands can also be restricted rather than removed, with allowed = ["explain"]. In every case, anything that runs is bounded by the privileges of the read-only role.

Natural-language queries depend on this. The feature generates SQL but cannot run it, so with commands off it has nothing to execute against and nothing to validate against. It also has a switch of its own, feature.ask_data_enabled, which your administrator can change at any time without a redeploy.

What network access does the collector need?

Section titled “What network access does the collector need?”

Outbound to otlp.dbgorilla.com and auth.dbgorilla.com on port 443, and no other destinations. Every connection the collector makes is outbound. It also listens on port 4318, for an intake it uses internally. It binds that port on every network interface of its container or pod and accepts data there without authentication, so do not publish the port, and on Kubernetes restrict it with a NetworkPolicy. Connections to your database use the TLS mode set in ssl_mode; verify-full is the recommended mode.

Can an agent in my IDE run arbitrary SQL against my database?

Section titled “Can an agent in my IDE run arbitrary SQL against my database?”

No. No raw SQL tool is exposed over MCP. Agents in your IDE cannot compose and execute arbitrary statements against your database through the integration.

Context only: query performance data, index usage, run frequency, topology and schema. An experiment on PostgreSQL can also return the optimizer statistics collect_statistics copied, which include values from your data. This removes the need for an engineer to run EXPLAIN ANALYZE against the primary on an agent’s behalf.

Can DBGorilla run EXPLAIN ANALYZE on production?

Section titled “Can DBGorilla run EXPLAIN ANALYZE on production?”

No. DBGorilla SaaS runs plan-only EXPLAIN through the collector, and only when you have enabled [commands]. EXPLAIN ANALYZE executes the query, and DBGorilla does not do that against your database.

MCP keys are scoped to a single user. Each user has one MCP key at a time; it is shown in the app, and regenerating or deleting it there revokes the previous one immediately. API keys are scoped to a single tenant. Both are encrypted at rest.

Yes. Clients that support it, such as Claude Desktop connectors, authenticate to the DBGorilla MCP server with OAuth rather than a bearer key. Other clients use a per-user bearer key.

This path is typically used by analysts to query production data that has not yet reached the warehouse. Three controls apply before a query runs, and a fourth disables the path entirely:

  • The generated query is shown for preview before it executes.
  • The collector runs the statement in a read-only session, so the database refuses anything that would modify or remove data. A second statement after a semicolon is rejected. On MySQL the read-only session does not stop DDL: a login granted CREATE can create a table this way, and it is kept. The grants on the login are what stop it.
  • Results are capped at 1,000 rows and a 30-second statement timeout, enforced at the collector.
  • Turning commands off everywhere in the collector config disables natural-language queries entirely.

It reviews pull requests and merge requests and posts comments. It does not apply changes to any database. Its findings are advisory.

A person can also ask for a pull request from a recommendation. On an open recommendation, every built-in role, Developer included, sees a Generate a PR button. The button stays disabled until a GitHub repository is linked to the database, as described in Link the repository to its database. The DBGorilla GitHub App opens the pull request on a new branch, as dbgorilla[bot], and leaves it for a person to merge. DBGorilla does not merge it. See Generate a PR.

Two tokens that belong to a bot user: one that posts the review, and a read-only one that reads the projects you turn on. Both are encrypted at rest. DBGorilla refuses a token that belongs to a person, so reviews are never posted as a teammate. The Maintainer token that adds a project’s webhook is used for that one request and never stored.

Webhook deliveries are verified before DBGorilla acts on them, with a signature on GitLab versions that sign webhooks and with a per-project secret otherwise. See Set up on GitLab.

Are recommendations tested before I see them?

Section titled “Are recommendations tested before I see them?”

Not always. When your collector allows collect_statistics, which every sample config does, a change can be tested in an experiment before it reaches you. The experiment runs on an empty, version-matched copy of your schema with your real optimizer statistics restored, in DBGorilla infrastructure, at our cost, not against your production instance. A recommendation that was tested cites the result in its evidence. Nothing requires the test: configuration and auditing changes, for example, reach you without one, and so does any recommendation an agent creates through the create_recommendation tool.

collect_statistics copies those statistics and no table rows. On PostgreSQL the statistics include each column’s most common values and histogram boundaries, which are values from your data. Without it, recommendations reach you without that test.

Metadata, schema and query patterns. Table rows are not sent to the model layer and are not written to logs. When an experiment runs on PostgreSQL, the model layer can also read the optimizer statistics collect_statistics copied, and those include each column’s most common values and histogram boundaries.

No. Customer data is never used to train models.

ControlStatus
Collector connects read-only, with permissions equivalent to a database monitoring agentEnforced today
DBGorilla has no direct network path to your database; every connection the collector makes is outboundEnforced today. The collector listens on port 4318 inside its container or pod, without authentication; keep that port unreachable.
Command execution can be turned off from the collector config, in your infrastructure, by setting enabled = false and leaving commands unset on every componentEnforced today
No raw SQL tool exposed over MCPEnforced today
EXPLAIN is plan-only; EXPLAIN ANALYZE is not runEnforced today
Generated statements run in a read-only session, one statement per requestEnforced today. On MySQL, DDL is stopped only by the login’s grants.
Query preview before a generated query executesEnforced today
Row-count limit (1,000) and 30-second timeout on natural-language query resultsEnforced at the collector. Not yet configurable in the application.
Recommendations validated in an experiment, on a copy of your schema with no rows, before they are surfacedAvailable when the collector allows collect_statistics, which every sample config does. Not enforced: a recommendation can be surfaced without an experiment.
MCP keys scoped to a single user; API keys scoped to a single tenant; both encrypted at restEnforced today
OAuth authentication for MCP clients that support itAvailable today
CI review bot posts review comments and applies no changes to any databaseEnforced today
Generate a PR on a recommendation, for a GitHub repository linked to the databaseAvailable today to every built-in role. Disabled until a repository is linked. The DBGorilla GitHub App opens the pull request and leaves it for a person to merge.
GitLab bot and read-only tokens encrypted at rest; tokens belonging to a person refused; Maintainer token never storedEnforced today
Audit log, with exportAvailable today on Business and Enterprise plans.
Single sign-on (SAML / OIDC)Available today on Business and Enterprise plans. SCIM and custom roles on Enterprise.

Security questions, vulnerability disclosures and requests for further review: security@dbgorilla.com.