Tailscale "fetch control key" connection reset by peer: the fix
Tailscale 1.104.1–1.104.6 can't reach its control server behind TLS-inspecting firewalls: connection reset or TLS handshake timeout. Update to 1.104.7.
If you run Tailscale 1.104.1 to 1.104.6 on an office, university or other filtered network, a firewall that inspects TLS is dropping the connection to Tailscale’s coordination server, because the newer Go release those versions are built with sends a TLS greeting the firewall does not understand. Update to 1.104.7 or later, released November 12, 2026, which Tailscale says fixes it. On Ubuntu or Debian: sudo apt-get update && sudo apt-get install tailscale, then check with tailscale version. If 1.104.7 and an older 1.102 both fail the same way, the network is blocking Tailscale outright, and the fix is the network admin allowing controlplane.tailscale.com on port 443.
The messages
In the tailscaled log the failure looks like this, from the report in tailscale/tailscale#21708:
fetch control key: Get "https://controlplane.tailscale.com/key?v=148": read tcp [local-address]:[port]->192.200.0.114:443: read: connection reset by peerCopiedThe number after ?v= is the client’s protocol version and differs between releases. Other reporters, on 1.104 behind the same kind of firewall, saw a timeout instead of a reset:
Do: net/http: TLS handshake timeoutCopiedThe apps show it as “Encrypted connection failed”: “Tailscale could not establish an encrypted connection with” the coordination server. Either way the device cannot fetch the server’s key, so it cannot sign in or get its list of peers, and uploads to log.tailscale.com fail beside it.
Why 1.104 broke on networks where 1.102 worked
Tailscale 1.104.x was built with Go 1.27, whose TLS client offers post-quantum options in its opening message, the ClientHello. Firewalls that inspect TLS read that message to decide what to let through, and some reject what they do not recognise. Reporters on Ubuntu 24.04, Debian, macOS, Android and Home Assistant OS found 1.102.x connecting on the same networks; one, behind Palo Alto firewalls on a corporate network, wrote “1.102.x works flawlessly”.
Which part of the message trips the firewall was not settled. A Tailscale maintainer put it down to the problem known as tldr.fail, where the larger ClientHello that offers ML-KEM is mishandled, and the issue’s title says the same. A user who replayed captured handshakes on a university network reported the opposite: “I don’t think ML-KEM is the trigger, at least here,” finding instead that the handshake passed once ML-DSA was removed from the signature_algorithms_cert list. The thread did not reconcile the two, and it does not need to: the fix covers both. Go 1.27.4 changed the ClientHello, and Tailscale’s changelog for 1.104.7 says “Go is updated to Go 1.27.4, to explicitly resolve an issue in TLS ClientHello where some middleboxes break connections to the Tailscale control plane.” The issue was closed on November 17 with “Resolved now that 1.104.7 is out.”
Update to 1.104.7 or later
Check what you have, and what the latest release is:
tailscale version
tailscale version --upstreamCopiedOn Linux, update with the package manager you installed from, as Tailscale’s update page lists it:
# Ubuntu, Debian
sudo apt-get update
sudo apt-get install tailscale
# Fedora, RHEL 8+, CentOS Stream 9
sudo dnf update tailscaleCopiedtailscale update does the same on some distributions. If you pinned 1.102.x to get through, as some reporters did, remove the pin, or the package manager will leave you there. For a server you will not be watching, tailscale set --auto-update keeps it current, so a fix like this one arrives without you.
Check the connection directly
The check Tailscale’s maintainers asked for in the issue tests the connection to the coordination server on its own:
tailscale debug ts2021 --verboseCopiedA Do: line ending in connection reset by peer or TLS handshake timeout is this failure. To tell a network from a client problem, try the same device from a phone’s hotspot, which the maintainers suggested because a mobile network should not be inspecting TLS: one reporter’s connection came back as soon as they left the corporate Wi-Fi. That tells you which network is filtering; it is a test, not a way to run.
When updating does not fix it
If 1.104.7, and an older release too, still fail on one network and work on another, the network is blocking the coordination server itself, not the new greeting. That was the finding in tailscale/tailscale#21981, on Ubuntu 24.04.5 with 1.104.2 and 1.102.4: the log ended in context deadline exceeded instead of a reset, and a handshake that named controlplane.tailscale.com timed out where one without the name went through. The maintainer closed it as “network filtering rather than a Tailscale defect”, since “tailscaled is not part of that exchange.” A plain request shows the same without Tailscale in the way:
curl -sv https://controlplane.tailscale.com/key?v=148 -o /dev/nullCopiedIf that fails too, the block is the network’s policy, and the people who run the network are the ones to fix it. What to ask for is on Tailscale’s firewall ports page: outbound HTTPS on port 443 to controlplane.tailscale.com, login.tailscale.com and log.tailscale.com, by name rather than by address (if address rules are unavoidable, 192.200.0.0/24 covers login and controlplane), plus TCP 443 and UDP 3478 to Tailscale’s relays. Where the network inspects TLS, that inspection is what Tailscale’s own page on the message names as a thing to check. Tunnelling round a block your employer or university set on purpose is not a fix; ask them.
Other things that stop a device reaching its tailnet look different. Another VPN that captures all traffic does it, as in Tailscale and NordVPN; a server whose /etc/resolv.conf was rewritten loses names, not the coordination server, as in the DNS config not ideal warning. Once a server is connected, locking it down so only your own devices reach it is in Tailscale SSH for an AI agent server.