You built it, it works. Now you get to own it. All of it.
The workflow went in fast, the business relies on it, and nobody minded until it touched customer data or an auditor asked. Then the questions change. Who patches this. Who is on call for it. Where are the logs. Who can change it, and how would we know.
None of those are questions about the tool. It is a fast and genuinely good way to build, and 2.0 tightened the defaults for every operator at once on 15 December 2025. They are questions about who operates it now, and self-hosted the answer is your team.
The patch queue is yours
n8n’s own documentation makes maintenance of a self-hosted instance your responsibility. Their security page scopes scans and penetration tests to n8n’s own production environment, and tells self-hosters to handle TLS and encryption at rest. Their Master Enterprise Terms make the customer solely responsible for the internal cost of implementing software updates, and disclaim liability for issues arising from outdated versions.
There is no version of Tray for you to be behind on.
129 CVE records in n8n’s own code, published between 15 December 2025 and 18 August 2026 (NVD, pulled 18 August 2026), counted by publication date, not affected version. The point is the frequency. The answer to any single CVE is that it was patched. Our n8n security tracker keeps the current high-and-critical subset live, with the version each one affects.
One of them is being exploited. CVE-2025-68613 was added to CISA’s Known Exploited Vulnerabilities catalogue on 11 March 2026. Listing carries a fourteen-day deadline for federal agencies. That catalogue holds roughly 1,670 entries, so a listing is not evidence n8n is unusually exposed. It is evidence that somebody patches on a deadline, and self-hosted that somebody is you.
2.0 did not end it. Five sandbox and validator bypasses have landed in the hardened paths themselves, between 18 January and 11 August 2026. On 22 May 2026 Singapore’s Cyber Security Agency reported that CVE-2026-44791 circumvented the fix for CVE-2026-42232, so patched instances were still vulnerable.
The architecture is now your problem
n8n publishes a figure of 200 executions per second, so the ceiling is not the argument. The architecture is, and their own docs are candid about it.
n8n does not restrict how much data a node fetches, which they say can cause errors when an execution runs out of memory. The documented remedies are provisioning more memory, or splitting workflows into smaller batches. The database can grow until it runs out of storage.
Retention, connection pressure during spikes and idempotency are yours to design and keep configured, indefinitely, alongside whatever your team was hired to build.
When you pay per execution, you are still paying for the servers
Free self-hosted Community is genuinely unmetered. Unlimited executions, no phone-home, a better offer than most vendors make.
What free excludes, per n8n’s own documentation: projects, SSO, log streaming, external secrets, environments, Git version control and sharing, with audit logging Enterprise only. Set that list beside a security questionnaire and the overlap is close to total. It also puts mitigations out of reach: Singapore’s guidance was to limit workflow editing to fully trusted users, and free Community has no Projects construct to do it with.
Above free, the meter runs on your own servers. €667 a month covers 40,000 executions, then €4,000 for each additional 300,000, and the license key must ping n8n’s server daily to stay active, reporting your production execution counts. Prices retrieved 18 August 2026. That plan is scoped to organizations under 100 employees, and below Enterprise the only support listed is the forum, with no published response-time target.
Where n8n wins
Speed of build, self-hosted control on infrastructure you already own, a large, active community with answers to most build questions, and an early, credible path to AI and agent workflows. If your team has the capacity to run a platform and is getting value from that flexibility, that is a coherent choice and it stays coherent.
What changes on Tray
- Patching is automatic. Tray patches and monitors the platform, and encrypts data in transit and at rest, so the next advisory is not your 2am page.
- The controls are already on. RBAC, SSO, audit logging and versioning are set up, kept working and evidenced by Tray, so a reviewer asks what the platform can show rather than what your team configured.
- Evidence a reviewer accepts. SOC 1 and SOC 2 Type 2 with annual penetration testing, HIPAA, GDPR and CCPA, and US, EU or APAC residency, with scope, subprocessors and signed DPAs on the trust center.
- Your workloads scale automatically. Nobody provisions workers before a spike and nobody re-architects after one.
- Policy on every agent and MCP call. Evaluated by the platform and kept as a platform record, so coverage does not depend on each builder adding the right node.
- Named support, on an SLA. Contractual response times and severity definitions, so “what happens when this breaks” is not a forum thread.
- You are not paying twice. The usage meter is not running on servers you are also buying, staffing and monitoring.
100% workflow execution uptime over the trailing 90 days, publicly measured on status.tray.ai, as of August 2026.
Working out where you actually stand
The production readiness checklist is 25 checks across security, compliance, agent governance, reliability, cost and support.