klever-go: REST API slow-header connection exhaustion via Gin Engine.RunGHSA-w4c6-7r69-w7j9

Summary

The Klever seednode REST API starts a Gin engine with Engine.Run(restAPIInterface). In Gin v1.9.1, Engine.Run calls Go's default http.ListenAndServe, which constructs an HTTP server without application-level ReadHeaderTimeout, ReadTimeout, or MaxHeaderBytes limits.

An unauthenticated client that can reach a REST listener bound with Klever's documented --rest-api-interface :8080 all-interface option can hold incomplete HTTP headers open indefinitely. In a local proof against the real cmd/seednode/api.Start path on v1.7.17, 120 slow-header connections caused 20/20 legitimate /log probes to fail with accept: too many open files. A fixed control using the same Gin router behind an explicit http.Server with ReadHeaderTimeout, ReadTimeout, and MaxHeaderBytes retained 0 slow connections and served 20/20 probes.

This report is distinct from the P2P advisories and from my direct-message goroutine report. This finding concerns Klever-owned HTTP REST startup code (cmd/seednode/api and network/api) using Gin Engine.Run without server-level header deadlines. It does not depend on MultiDataInterceptor, Batch.Decompress, libp2p, malformed P2P messages, or direct-message goroutine spawning.

Details

Seednode REST API, latest release v1.7.17:

• cmd/seednode/api/api.go:17 defines Start(restAPIInterface, marshalizer).
• cmd/seednode/api/api.go:18 creates ws := gin.Default().
• cmd/seednode/api/api.go:23 returns ws.Run(restAPIInterface).
• cmd/seednode/CLI.md:23 documents --rest-api-interface; it says :8080 binds all interfaces and off disables the API.

Node REST API, latest release v1.7.17:

• network/api/api.go:79 creates ws = gin.Default().
• network/api/api.go:98 returns ws.Run(kleverFacade.RestAPIInterface()).
• cmd/node/main.go:147-150 documents the same --rest-api-interface flag and says :8080 binds all interfaces.
• docker/README.md:56-61 and docker/README.md:67-70 publish host port 8080 for full-node and validator Docker examples.
• README.md:264-268 documents that the node exposes a REST API for blockchain queries and operations.

The seednode REST API source is byte-identical across v1.7.14 through v1.7.17; the captured runtime PoC was executed on v1.7.17.

Current develop commit 10bcfd50 remains affected:

• network/api/api.go:98 still returns ws.Run(kleverFacade.RestAPIInterface()).
• cmd/seednode/api/api.go:59 still returns ws.Run(restAPIInterface).

Gin v1.9.1 implements Engine.Run as:

func (engine *Engine) Run(addr ...string) (err error) {
    address := resolveAddress(addr)
    err = http.ListenAndServe(address, engine.Handler())
    return
}

In my source sweep, I did not find a production http.Server{ReadHeaderTimeout: ...} wrapper for either REST start path. The only ReadHeaderTimeout hit I found in the repository was a test helper under network/api/websocket/routes_test.go.

PoC

GitHub Private Vulnerability Reporting does not appear to allow file attachments in this form, so I am including the reproduction command and captured output inline. I can paste the full 254-line Go test patch in a reply immediately if useful.

The test starts two local child servers:

1. Vulnerable: the real cmd/seednode/api.Start path.
2. Fixed control: the same Gin router served through http.Server{ReadHeaderTimeout: 250ms, ReadTimeout: 250ms, MaxHeaderBytes: 4096}.

Reproduction from a clean checkout:

git clone https://github.com/klever-io/klever-go
cd klever-go
git checkout v1.7.17

# Apply the PoC patch to cmd/seednode/api.
# I can provide the full patch in this advisory thread.

go test ./cmd/seednode/api -run TestPoC_SeednodeAPISlowlorisDifferential -count=1 -v -timeout 60s

Captured output on v1.7.17:

POC_RESULT mode=vulnerable slow_connections_opened=120 slow_connections_still_open=111 legitimate_probe_ok=0 legitimate_probe_fail=20
POC_RESULT mode=fixed slow_connections_opened=120 slow_connections_still_open=0 legitimate_probe_ok=20 legitimate_probe_fail=0

The vulnerable server also logs repeated accept failures:

http: Accept error: accept tcp 127.0.0.1:56415: accept: too many open files; retrying in 1s

Impact

For an externally reachable Klever REST listener, a single unauthenticated client can retain many server-side connections by never completing HTTP headers. Because the Go server has no read-header deadline, those connections persist until the client closes them or an external proxy/firewall intervenes.

The direct result is REST API unavailability for legitimate clients. The local proof demonstrates this as 0/20 legitimate /log probes succeeding while the vulnerable server is saturated, versus 20/20 succeeding with the fixed server wrapper.

I am not claiming default public internet exposure. The default bind is localhost:8080. The affected condition is a REST API listener exposed through Klever's documented all-interface bind or Docker port-publish deployment shape.

This maps to the SECURITY.md High category: "Denial of Service affecting network availability." If Klever treats externally reachable REST API unavailability as non-critical because the default bind is localhost, the conservative classification is Medium under "Performance degradation attacks" / "Non-critical DoS vectors."

All testing was local loopback only. I did not contact Klever mainnet, public testnet, hosted RPCs, explorers, or third-party production infrastructure.

Suggested fix:

Start both REST APIs through explicit http.Server values instead of Engine.Run, for example:

srv := &http.Server{
    Addr:              restAPIInterface,
    Handler:           ws.Handler(),
    ReadHeaderTimeout: 5 * time.Second,
    ReadTimeout:       10 * time.Second,
    WriteTimeout:      30 * time.Second,
    IdleTimeout:       120 * time.Second,
    MaxHeaderBytes:    32 << 10,
}
return srv.ListenAndServe()

Apply the same pattern to:

• cmd/seednode/api.Start
• network/api.Start

If Klever expects deployments to expose the REST API through a reverse proxy, I still recommend setting server-level limits in the application. That keeps the binary safe when operators use the documented direct bind or Docker port-publish path.