GHSA-pc2w-4mq8-32qw: Dynatrace MCP: notebook tool skips approval gate

### Summary A missing human-approval gate on the `create_dynatrace_notebook` tool allows a caller to create persistent tenant-visible documents containing arbitrary content (including embedded DQL that other users execute when opening the notebook) without operator consent. ### Details `dynatrace-mcp-server` registers six write tools: `send_slack_message`, `send_email`, `send_event`, `create_workflow_for_notification`, `make_workflow_public`, and `create_dynatrace_notebook`. Five of these call `requestHumanApproval()` before executing the side-effect, which elicits the operator's explicit consent through the MCP elicitation protocol. The CHANGELOG explicitly states these approval gates were added "to ensure user consent and prevent unintended actions." `create_dynatrace_notebook` does not call `requestHumanApproval()`. The tool was introduced in a separate release from the approval-gate retrofit and was left ungated. As a result, a caller can create persistent, tenant-visible Dynatrace notebooks containing arbitrary content with no operator confirmation. Notebooks can include embedded DQL queries that later execute under the permissions of any tenant user who opens them. The vulnerable code is in `src/index.ts`, lines 1563-1601: ```typescript tool( 'create_dynatrace_notebook', 'Create Dynatrace Notebook', 'Create a new notebook in the Dynatrace platform ...', { name: z.string().describe(/* ... */), description: z.string().optional().describe(/* ... */), content: z.array(z.object({ type: z.enum(['dql', 'markdown']), text: z.string(), })).describe(/* ... */), }, { readOnlyHint: false, }, async ({ name, content, description }) => { const dtClient = await createAuthenticatedHttpClient(allRequiredScopes); const data = await createDynatraceNotebook(dtClient, name, content, description); // No requestHumanApproval() call. // No destructiveHint annotation. // Scopes requested are allRequiredScopes (the broadest possible set) // even though only document:documents:write is needed. return data ? `Document created successfully: ${dtEnvironment}/ui/apps/dynatrace.notebooks/notebook/${data.id}` : 'document creation failed'; }, ); ``` By comparison, every other write tool calls `requestHumanApproval()` as its first action: - `send_email` (line 1274): `const approved = await requestHumanApproval(...);` - `send_slack_message` (line 652): `const approved = await requestHumanApproval(...);` - `send_event` (line 1367): `const approved = await requestHumanApproval(...);` - `create_workflow_for_notification` (line 1069): `const approved = await requestHumanApproval(...);` - `make_workflow_public` (line 1118): `const approved = await requestHumanApproval(...);` ### PoC Tested end-to-end against a real Dynatrace tenant. The MCP server was started in HTTP mode with the operator's Platform Token: ```bash export DT_ENVIRONMENT=https://<tenant>.apps.dynatrace.com export DT_PLATFORM_TOKEN=dt0s16.... npx -y @dynatrace-oss/dynatrace-mcp-server@1.8.5 --http --port 3000 ``` A single unauthenticated POST from a process with no Dynatrace credentials of its own: ```bash curl -sS http://127.0.0.1:3000/ \ -H 'Content-Type: application/json' \ -H 'Accept: application/json, text/event-stream' \ -d '{ "jsonrpc": "2.0", "id": 1, "method": "tools/call", "params": { "name": "create_dynatrace_notebook", "arguments": { "name": "Unapproved notebook", "description": "Created without operator approval.", "content": [ {"type": "markdown", "text": "# This notebook was created without approval"}, {"type": "dql", "text": "fetch logs | limit 10"} ] } } }' ``` The response: ``` event: message data: {"result":{"content":[{"type":"text","text":"Document created successfully: https://<tenant>.apps.dynatrace.com/ui/apps/dynatrace.notebooks/notebook/<uuid>"}]}, "jsonrpc":"2.0","id":1} ``` The notebook is visible in the operator's Notebooks app within seconds. No elicitation prompt was sent to any operator client. By contrast, the same harness invoking `send_email` or `send_slack_message` returns: ``` "Operation cancelled: Human approval was not granted for sending this email." ``` The gate logic exists and is wired up for all other write tools - it is simply missing on `create_dynatrace_notebook`. ### Impact - A caller (an unauthenticated network attacker over the HTTP transport when authentication is missing, or a prompt-injected LLM in stdio mode) can silently create persistent tenant-visible documents. - The notebook content is attacker-controlled. Embedded DQL queries execute under the permissions of any tenant user who later opens the notebook - a stored-DQL pattern that crosses identity boundaries within the tenant.

An attacker with network access to an HTTP-mode deployment of the MCP server (or, in stdio mode, an attacker who successfully prompt-injects the LLM the server is wired into) sends a single `tools/call` request for `create_dynatrace_notebook` with attacker-authored markdown and an embedded DQL query. Because this tool never calls `requestHumanApproval()`, the notebook is created and made tenant-visible within seconds, with no elicitation prompt reaching the operator — unlike a call to `send_email` or `send_slack_message`, which would be blocked. A legitimate, higher-privileged tenant user later opens the notebook out of routine curiosity; the embedded DQL executes under their session, letting the attacker read data or trigger actions outside their own original access level, all while looking like a normal, self-created document.