Dynatrace MCP Server: workflow template injection, DQL injection and a write tool without human approval (CVE-2026-123809, CVE-2026-123810, CVE-2026-123812), fixed by 2.1.1
Three more CVEs for the official Dynatrace MCP Server (npm @dynatrace-oss/dynatrace-mcp-server) show how an MCP caller, such as an agent following injected instructions, could go beyond what the tools promise: plant Jinja2 expressions in a persistent notification workflow that leak event data (fixed 2.0.0), inject DQL through parameters typed as identifiers (fixed 2.1.1), and create tenant-visible notebooks without the human-approval prompt the other write tools use (fixed 1.8.7). Upgrade to 2.1.1 or later.
What it is
CVE-2026-123809 / GHSA-xrmj-5g4g-8987 (fixed 2.0.0): create_workflow_for_notification puts the caller's teamName, problemType and channel into a Dynatrace Workflow. Workflows evaluate {{ ... }} Jinja2 expressions at runtime, so a value like {{ event() }} sends full event data to the Slack channel every time the workflow runs, and the workflow stays in the tenant after the MCP session ends.
CVE-2026-123810 / GHSA-pqh8-p93p-2rx7 (fixed 2.1.1): several read tools (find-monitored-entity-by-name, list-problems, list-vulnerabilities and others) paste caller values documented as identifiers or timeframes into DQL without escaping, so a caller can break out of the string, add pipeline stages and comment out the rest, getting around the tools' field, time-window and display limits.
CVE-2026-123812 / GHSA-pc2w-4mq8-32qw (fixed 1.8.7, low): create_dynatrace_notebook writes a persistent, tenant-visible notebook (which can embed DQL other users run when they open it) without calling requestHumanApproval(), unlike the other five write tools.
All three were reported by yotampe-pluto. A separate high-severity issue, CVE-2026-123811 (unauthenticated tools/call over --http before 2.0.0), already has its own VibeVulns entry.
How to check
- Run npm ls @dynatrace-oss/dynatrace-mcp-server (or check the version pinned in your MCP client config / npx command). Anything below 2.1.1 has at least CVE-2026-123810.
- In Dynatrace, list workflows whose title starts with "[MCP POC] Notify team" and look for {{ ... }} expressions in their inputs; also review notebooks created by the MCP server's token.
How to fix
- Upgrade to @dynatrace-oss/dynatrace-mcp-server 2.1.1 or later (pin the version in npx configs, e.g. @2.1.1).
- Delete unexpected MCP-created workflows and notebooks, and scope the platform token the server uses to the least it needs.
npm ls -g @dynatrace-oss/dynatrace-mcp-server; npm view @dynatrace-oss/dynatrace-mcp-server version
grep -rn 'dynatrace-mcp-server' ~/.cursor ~/.config ~/Library/Application\ Support/Claude 2>/dev/null Sources
- criticalMalicious npm package @guicai777/mcp-doctor posed as an MCP diagnostic tool and dropped a hidden Windows executable, with a README telling users to run mshta (MAL-2026-20295)
- criticalCVE-2026-77359: Jupyter MCP Server's /api/connect and /api/stop routes skipped the token and Origin checks, so a website or network caller could repoint it to an attacker's Jupyter backend (fixed 1.0.3)
- criticalMalicious npm packages posing as AI SDKs and MCP servers: character-ai-sdk, char-ai-sdk, @mibbp/agent-observer, mrnima-openai-chatgpt-bot and @fdcasj_dfew/md-doc-mcp
- criticalCVE-2026-76807: Doist's official Todoist MCP HTTP server let any website use DNS rebinding to read, create and delete your Todoist data (fixed 10.3.3)