feat(security): add a --security "anonymous=..." posture, defaulting to full

dgraph/cmd/alpha/http.go

@@ -240,8 +240,7 @@ func queryHandler(w http.ResponseWriter, r *http.Request) {
 	}
 
 	ctx := context.WithValue(r.Context(), query.DebugKey, isDebugMode)
-	ctx = x.AttachAccessJwt(ctx, r)
-	ctx = x.AttachRemoteIP(ctx, r)
+	ctx = x.AttachRequestIdentity(ctx, r)
 
 	if queryTimeout != 0 {
 		var cancel context.CancelFunc
@@ -467,7 +466,7 @@ func mutationHandler(w http.ResponseWriter, r *http.Request) {
 	req.Hash = hash
 	req.CommitNow = commitNow
 
-	ctx := x.AttachAccessJwt(context.Background(), r)
+	ctx := x.AttachRequestIdentity(context.Background(), r)
 	resp, err := (&edgraph.Server{}).QueryNoGrpc(ctx, req)
 	if err != nil {
 		x.SetStatusWithData(w, x.ErrorInvalidRequest, err.Error())
@@ -530,7 +529,7 @@ func commitHandler(w http.ResponseWriter, r *http.Request) {
 		return
 	}
 
-	ctx := x.AttachAccessJwt(context.Background(), r)
+	ctx := x.AttachRequestIdentity(context.Background(), r)
 	var response map[string]interface{}
 	if abort {
 		response, err = handleAbort(ctx, startTs, hash)

dgraph/cmd/alpha/run.go

@@ -173,7 +173,19 @@ they form a Raft group and provide synchronous replication.
 			"A comma separated list of IP addresses, IP ranges, CIDR blocks, or hostnames you wish "+
 				"to whitelist for performing admin actions (i.e., --security "+
 				`"whitelist=144.142.126.254,127.0.0.1:127.0.0.3,192.168.0.0/16,host.docker.`+
-				`internal").`).
+				`internal"). Empty by default, which admits loopback only. This is a network `+
+				"location check, NOT authentication: every address in the range can run "+
+				"privileged operations without a credential unless you also set token= or enable "+
+				"ACL.").
+		Flag("anonymous",
+			"[full, data, none] What a caller that presents no verified credential may do. "+
+				"full (default) leaves authorization to whatever the whitelist, token, and ACL "+
+				"settings decide, which is the behavior of every earlier release. data allows "+
+				"queries, mutations, commits, and login while denying every administrative "+
+				"operation, including schema changes and drops, regardless of the whitelist. "+
+				"none additionally denies queries, "+
+				"mutations, and commits, leaving only login and the health endpoints. data and "+
+				"none require token= or ACL, otherwise no caller can ever be identified.").
 		String())
 
 	flag.String("limit", worker.LimitDefaults, z.NewSuperFlagHelp(worker.LimitDefaults).
@@ -310,7 +322,10 @@ func healthCheck(w http.ResponseWriter, r *http.Request) {
 		w.Header().Set("Content-Type", "application/json")
 		w.WriteHeader(http.StatusOK)
 
-		ctx := x.AttachAccessJwt(context.Background(), r)
+		// Full identity, not just the access JWT: Health(all) is a capability check,
+		// and under a closed --security "anonymous=..." posture a caller presenting
+		// only the --security token would otherwise arrive unidentified.
+		ctx := x.AttachRequestIdentity(context.Background(), r)
 		var resp *api.Response
 		if resp, err = (&edgraph.Server{}).Health(ctx, true); err != nil {
 			x.SetStatus(w, x.Error, err.Error())
@@ -355,8 +370,9 @@ func stateHandler(w http.ResponseWriter, r *http.Request) {
 	x.AddCorsHeaders(w)
 	w.Header().Set("Content-Type", "application/json")
 
-	ctx := context.Background()
-	ctx = x.AttachAccessJwt(ctx, r)
+	// Full identity, not just the access JWT: State is a capability check. See the
+	// note on the /health?all branch above.
+	ctx := x.AttachRequestIdentity(context.Background(), r)
 
 	var aResp *api.Response
 	if aResp, err = (&edgraph.Server{}).State(ctx); err != nil {
@@ -652,6 +668,8 @@ func run() {
 		FromSuperFlag(Alpha.Conf.GetString("badger"))
 	security := z.NewSuperFlag(Alpha.Conf.GetString("security")).MergeAndCheckDefault(
 		worker.SecurityDefaults)
+	anonymous, err := x.ParseAnonymousPosture(security.GetString("anonymous"))
+	x.Check(err)
 	conf := audit.GetAuditConf(Alpha.Conf.GetString("audit"))
 
 	x.Config.Limit = z.NewSuperFlag(Alpha.Conf.GetString("limit")).MergeAndCheckDefault(
@@ -720,6 +738,7 @@ func run() {
 		AbortOlderThan:      abortDur,
 		StartTime:           startTime,
 		Security:            security,
+		Anonymous:           anonymous,
 		TLSClientConfig:     tlsClientConf,
 		TLSServerConfig:     tlsServerConf,
 		AclJwtAlg:           keys.AclJwtAlg,
@@ -729,18 +748,37 @@ func run() {
 	}
 	x.WorkerConfig.Parse(Alpha.Conf)
 
+	// The built-in authenticator: ACL's access JWT, plus the --security auth token as
+	// an identity rather than a boolean check. Installed before ConfigureIdentity so a
+	// deployment can compose with or replace it; see edgraph.PresharedAuthenticator.
+	x.SetAuthenticator(edgraph.PresharedAuthenticator())
+
 	// Install deployment-specific authentication and authorization now: the config
 	// is parsed, and nothing is serving yet. A misconfiguration here is fatal, which
 	// is why it runs before any listener rather than lazily on the first request.
 	ConfigureIdentity()
 
+	// The --security "anonymous=..." floor goes on last, so that it wraps whatever
+	// policy ConfigureIdentity installed rather than being replaced by it. Under the
+	// shipped anonymous=full this installs nothing.
+	edgraph.EnforceAnonymousPosture(anonymous)
+
 	// Set the directory for temporary buffers.
 	z.SetTmpDir(x.WorkerConfig.TmpDir)
 
 	x.WorkerConfig.EncryptionKey = keys.EncKey
 
 	setupCustomTokenizers()
 	x.Config.PortOffset = Alpha.Conf.GetInt("port_offset")
+
+	// Deliberately this late. It has to follow ConfigureIdentity, because a
+	// deployment authenticator changes who can be identified, and it has to follow
+	// the port offset, because the warning names the HTTP port.
+	builtinIdentity := x.AuthenticatorName() == edgraph.PresharedAuthenticator().Name()
+	for _, msg := range securityWarnings(anonymous, security.GetString("whitelist"), ips,
+		opts.AuthToken, keys.AclSecretKey != nil, builtinIdentity, httpPort()) {
+		glog.Warning(msg)
+	}
 	x.Config.LimitMutationsNquad = int(x.Config.Limit.GetInt64("mutations-nquad"))
 	x.Config.LimitQueryEdge = x.Config.Limit.GetUint64("query-edge")
 	x.Config.BlockClusterWideDrop = x.Config.Limit.GetBool("disallow-drop")

dgraph/cmd/alpha/security_posture.go

@@ -0,0 +1,87 @@
+/*
+ * SPDX-FileCopyrightText: © 2017-2026 Istari Digital, Inc.
+ * SPDX-License-Identifier: Apache-2.0
+ */
+
+package alpha
+
+import (
+	"fmt"
+
+	"github.com/dgraph-io/dgraph/v25/x"
+)
+
+// securityWarnings returns the startup warnings for a --security configuration
+// whose parts do not add up, or nil when they do.
+//
+// Three controls gate Alpha's privileged operations -- the whitelist, the auth
+// token, and ACL -- and each one passes when its own feature is unconfigured. The
+// protection an operator gets is whatever they configured rather than the union of
+// the three, and the two combinations below are the ones where that produces an
+// outcome they almost certainly did not intend.
+//
+// It is pure and returns strings rather than logging, so the combinations can be
+// pinned by a table test.
+//
+// builtinIdentity is false when a deployment installed its own authenticator via
+// ConfigureIdentity. Such an authenticator can produce identities from credentials
+// this function knows nothing about -- an external JWT, say -- so whether a closed
+// posture can be administered is no longer something it can judge.
+func securityWarnings(posture x.AnonymousPosture, whitelist string, ips []x.IPRange,
+	authToken string, aclEnabled, builtinIdentity bool, httpPort int) []string {
+
+	hasCredential := aclEnabled || authToken != ""
+	var out []string
+
+	// The shipped default is an empty whitelist, which admits loopback only, so the
+	// admin plane does not leave the host without an operator widening it. This
+	// warns at exactly that point: whitelisting answers where a request came from
+	// and has no credential in it, so a widened range with nothing else configured
+	// means every address inside it can run privileged operations anonymously.
+	if posture == x.AnonymousFull && admitsNonLoopback(ips) && !hasCredential {
+		out = append(out, fmt.Sprintf(
+			`SECURITY: --security "whitelist=%s" admits non-loopback callers, but neither ACL nor `+
+				`an admin token is configured. Privileged operations (backup, restore, export, `+
+				`shutdown, removeNode, moveTablet, assign, draining, config, namespace create and `+
+				`drop) are reachable from that range WITHOUT ANY CREDENTIAL: anyone who can reach `+
+				`port %d can read the whole database via backup or export, restore over it, or shut `+
+				`the cluster down. Set --security "token=..." or enable ACL, and narrow the `+
+				`whitelist to the addresses that actually administer this cluster. `+
+				`--security "anonymous=data" additionally denies every administrative operation to `+
+				`a caller that presents no credential.`, whitelist, httpPort))
+	}
+
+	// A closed posture with nothing that can produce an identity. Every capability
+	// check will deny, including the operator's own, so the cluster cannot be
+	// administered at all. This is a misconfiguration rather than a hardening, and
+	// it is worth saying so loudly at boot instead of letting it surface as a
+	// permission error during an incident.
+	if posture.RequiresIdentityForCapability() && !hasCredential && builtinIdentity {
+		out = append(out, fmt.Sprintf(
+			`SECURITY: --security "anonymous=%s" requires an identified caller, but neither ACL nor `+
+				`an admin token is configured, so no request can ever be identified. Every `+
+				`administrative operation will be denied, including from loopback, and this cluster `+
+				`cannot be administered. Set --security "token=..." or enable ACL.`, posture))
+	}
+
+	return out
+}
+
+// admitsNonLoopback reports whether any whitelist range admits an address outside
+// loopback. Loopback is always admitted regardless of the whitelist, so a range
+// that covers only loopback -- whitelist=127.0.0.1, or the documented "localhost
+// only" example -- does not take the admin plane off the host and is not worth a
+// warning.
+//
+// A range is loopback-only when both of its ends are loopback. That is exact
+// rather than approximate: IPv4 loopback is the contiguous block 127.0.0.0/8, and
+// IPv6 loopback is the single address ::1, so a range whose ends both fall inside
+// one of those cannot contain anything else.
+func admitsNonLoopback(ips []x.IPRange) bool {
+	for _, r := range ips {
+		if !r.Lower.IsLoopback() || !r.Upper.IsLoopback() {
+			return true
+		}
+	}
+	return false
+}