GHSA-26gq-p25f-99cp

## Summary

An integer-overflow vulnerability in the frp server's optional SSH Tunnel Gateway lets any unauthenticated remote attacker crash the entire `frps` process with a single five-byte message. When the gateway parses an SSH `exec` channel request in `pkg/ssh/server.go`, it adds a small constant to a four-byte length value taken directly from the request. Because that length is fully attacker-controlled, a value of `0xFFFFFFFF` makes the addition wrap around to a tiny number, defeating the only bounds check and forcing an out-of-range slice. Go raises a panic that nothing recovers, so the whole server exits. In the default gateway mode SSH clients are not authenticated and the request is handled before any frp token is checked, so the crash is reachable pre-authentication. It carries no state and is trivially repeatable, turning one crash into a sustained outage that drops every tunnel for every user.

## Details

The `exec` request payload is a four-byte big-endian length followed by that many command bytes, and the length field is fully attacker-controlled. The gateway computes the end of the command as `4 + length`. The constant `4` takes the `uint32` type of the length field, so `4 + 0xFFFFFFFF` wraps modulo 2^32 to `3`. The only guard compares the real payload size against this wrapped value, so a five-byte payload passes the test `5 < 3`, and the slice runs from index 4 to index 3:

```go // pkg/ssh/server.go:315-319 end := 4 + binary.BigEndian.Uint32(req.Payload[:4]) // 4 + 0xFFFFFFFF wraps to 3 if len(req.Payload) < int(end) { // 5 < 3 is false, so it passes continue } extraPayload := string(req.Payload[4:end]) // payload[4:3] -> panic ```

The handler runs in a bare goroutine and no function in the call chain installs a `recover`, so the panic unwinds to the top of the goroutine and terminates the whole process rather than the single connection:

```go // pkg/ssh/server.go:226 go s.handleNewChannel(newChannel, extraPayloadCh) // no recover anywhere in the chain ```

The sink is reachable without credentials in the default configuration. When no authorized-keys file is set, the gateway disables SSH client authentication, which is the default and documented mode where users authenticate through the frp token embedded in the SSH command. The `exec` request is processed during the channel phase, before that token is validated, so an unauthenticated peer reaches the sink:

```go // pkg/ssh/gateway.go:74 sshConfig.NoClientAuth = cfg.AuthorizedKeysFile == "" // default: no client auth ```

Affected product: frp (fatedier), `frps` server, SSH Tunnel Gateway Affected versions: v0.53.0 (when the gateway was introduced) through v0.70.0, confirmed against a v0.70.0 build

## PoC

The SSH Tunnel Gateway must be enabled with no authorized-keys file, which is the default mode of any deployment that turns the gateway on. No credentials and no knowledge of the target are required.

Enable the gateway on a stock v0.70.0 build:

```toml bindPort = 7000

[sshTunnelGateway] bindPort = 2200 ```

As an unauthenticated SSH client, open a session channel and send one `exec` request whose four-byte length prefix is `0xFFFFFFFF` followed by a single byte, giving the five-byte payload `ff ff ff ff 41`. The server immediately panics and exits:

```text panic: runtime error: slice bounds out of range [4:3] ...github.com/fatedier/frp/pkg/ssh.(*TunnelServer).handleNewChannel .../frp/pkg/ssh/server.go:319 ```

The shell prompt returns in the server terminal, confirming the whole process has terminated and every other client tunnel is dropped at the same instant. A well-formed request of the same length (prefix `1`) leaves the server running, proving the crash is caused by the overflow value and not by the `exec` request itself. The attacker can re-send the payload after any restart to hold the server down indefinitely.

## Impact

This is an unauthenticated remote denial of service (CWE-190) leading to an out-of-range array index (CWE-129) and an unhandled process crash (CWE-755). A single five-byte, credential-free message kills the entire `frps` process. Because one `frps` commonly multiplexes the tunnels of many clients, the crash drops every active tunnel for every tenant at once and new logins are refused until the process is restarted, and re-sending the payload after each restart keeps the service down indefinitely (CWE-400). There is no confidentiality or integrity impact and no code execution.

This issue only affects frps instances with the SSH tunnel gateway explicitly enabled. The SSH tunnel gateway is disabled by default and is not commonly used in typical frps deployments. Default frps configurations are not affected.