diff --git a/ai/cmd/agent/main.go b/ai/cmd/agent/main.go
index b275567..1efaee0 100644
--- a/ai/cmd/agent/main.go
+++ b/ai/cmd/agent/main.go
@@ -122,9 +122,17 @@ func main() {
 	// external-* model ids (reference runs, 2026-10-16): each loop turn is served to an outside client over
 	// GET /external/next and answered through POST /external/reply (internal/external). The diagnosis deadline is
 	// EXTERNAL_DIAGNOSE_TIMEOUT_S (the harness waits a little longer for external-* models), one turn at most
-	// EXTERNAL_TURN_TIMEOUT_S. The endpoints are reachable only inside the VPC / over the IAP SSH tunnel.
-	ext := external.New(time.Duration(envInt("EXTERNAL_TURN_TIMEOUT_S", 1200)) * time.Second)
-	ag.SetExternal(ext)
+	// EXTERNAL_TURN_TIMEOUT_S. Off unless EXTERNAL_ADDR is set; the endpoints live on that separate listener, which
+	// the VM compose file maps to 127.0.0.1 only (reachable through the IAP SSH tunnel, never from the VPC): a reply
+	// is a diagnosis and, under auto-approval, an executed action.
+	var ext *external.Responder
+	if addr := env("EXTERNAL_ADDR", ""); addr != "" {
+		ext = external.New(time.Duration(envInt("EXTERNAL_TURN_TIMEOUT_S", 1200)) * time.Second)
+		ag.SetExternal(ext)
+		esrv := &http.Server{Addr: addr, Handler: logging(ext.Handler()), ReadHeaderTimeout: 10 * time.Second}
+		go func() { log.Fatal(esrv.ListenAndServe()) }()
+		log.Printf("external responder: listening on %s", addr)
+	}
 	externalDiagnoseTimeout := time.Duration(envInt("EXTERNAL_DIAGNOSE_TIMEOUT_S", 1500)) * time.Second
 	jevMode := env("JEV_MODE", "off") // off | shadow | enforce (shadow records Jev answers next to the diagnosis; enforce also applies the two review rules)
 	var jc *jev.Client
@@ -166,7 +174,11 @@ func main() {
 		if f := anth.Federation(); f != nil {
 			out["anthropic_federation"] = f.Probe(c, anth.BaseURL) // one real token exchange; the Console history page shows the deny reason on failure
 		}
-		out["external_responder"] = ext.Status()
+		if ext != nil {
+			out["external_responder"] = ext.Status()
+		} else {
+			out["external_responder"] = "off"
+		}
 		out["jev_mode"] = jevMode
 		if jc != nil {
 			out["jev_key_present"] = jc.KeyPresent()
@@ -222,8 +234,6 @@ func main() {
 		ag.SaveShadow(context.Background(), req, res.SessionID, res, tri, rev)
 		writeJSON(w, 200, res)
 	})
-	mux.HandleFunc("GET /external/next", ext.Next)
-	mux.HandleFunc("POST /external/reply", ext.ReplyHandler)
 	srv := &http.Server{Addr: env("AGENT_ADDR", ":8080"), Handler: logging(mux), ReadHeaderTimeout: 10 * time.Second}
 	log.Fatal(srv.ListenAndServe())
 }
diff --git a/ai/internal/external/responder.go b/ai/internal/external/responder.go
index 8238c8f..ab0cc37 100644
--- a/ai/internal/external/responder.go
+++ b/ai/internal/external/responder.go
@@ -3,10 +3,18 @@
 // GET /external/next (long poll); the client answers with POST /external/reply. Used for reference measurements with
 // a model that has no API path (2026-10-16: Claude Mythos 5.1 driven by a Claude Code subagent). Token usage is not
 // known for such turns and is reported as zero; reports mark those runs as reference records.
+//
+// A reply becomes a diagnosis and, with auto-approval, an executed action, so whoever can answer a turn controls the
+// lab: the endpoints are served on their own listener (EXTERNAL_ADDR, off by default; the VM maps it to 127.0.0.1 so
+// only the IAP SSH tunnel reaches it), turn ids are random, and every reply must carry the per-turn reply_token that
+// only the client who took the turn received.
 package external
 
 import (
 	"context"
+	"crypto/rand"
+	"crypto/subtle"
+	"encoding/hex"
 	"encoding/json"
 	"fmt"
 	"net/http"
@@ -24,7 +32,8 @@ const maxConvCached = 64
 // on the first turn, so it is carried along), the tool definitions and whether the final submit_diagnosis is forced.
 type Turn struct {
 	TurnID          string        `json:"turn_id"`
-	Session         string        `json:"session"` // first response id of the conversation ("" on the first turn)
+	ReplyToken      string        `json:"reply_token"` // random, handed out with the turn, required by /external/reply
+	Session         string        `json:"session"`     // first response id of the conversation ("" on the first turn)
 	Model           string        `json:"model"`
 	Turn            int           `json:"turn"` // 1-based index within the conversation
 	System          string        `json:"system"`
@@ -38,9 +47,10 @@ type Turn struct {
 
 // Reply is the client's answer: function calls (name + arguments) and/or assistant text.
 type Reply struct {
-	TurnID string      `json:"turn_id"`
-	Output []replyItem `json:"output"`
-	Error  string      `json:"error,omitempty"` // client gave up on this turn
+	TurnID     string      `json:"turn_id"`
+	ReplyToken string      `json:"reply_token"`
+	Output     []replyItem `json:"output"`
+	Error      string      `json:"error,omitempty"` // client gave up on this turn
 }
 
 type replyItem struct {
@@ -115,11 +125,11 @@ func (r *Responder) CreateResponse(ctx context.Context, req openai.Request) (*op
 	if m, ok := req.ToolChoice.(map[string]any); ok && m["type"] == "function" {
 		forced = true
 	}
+	id, tok := "ext_"+randHex(8), randHex(16)
 	r.mu.Lock()
 	r.seq++
-	id := fmt.Sprintf("ext_%d", r.seq)
 	now := time.Now()
-	p := &pending{turn: Turn{TurnID: id, Session: hist.first, Model: req.Model, Turn: hist.turn + 1, System: hist.system, Input: hist.items, Tools: req.Tools,
+	p := &pending{turn: Turn{TurnID: id, ReplyToken: tok, Session: hist.first, Model: req.Model, Turn: hist.turn + 1, System: hist.system, Input: hist.items, Tools: req.Tools,
 		ForceSubmit: forced, MaxOutputTokens: req.MaxOutputTokens, IssuedAt: now, DeadlineAt: now.Add(r.TurnTimeout)}, reply: make(chan Reply, 1)}
 	r.queue = append(r.queue, p)
 	r.lastTurn = now
@@ -269,6 +279,11 @@ func (r *Responder) ReplyHandler(w http.ResponseWriter, req *http.Request) {
 	}
 	r.mu.Lock()
 	p, ok := r.taken[rep.TurnID]
+	if ok && subtle.ConstantTimeCompare([]byte(p.turn.ReplyToken), []byte(rep.ReplyToken)) != 1 {
+		r.mu.Unlock()
+		writeJSON(w, 403, map[string]string{"error": "reply_token does not match turn " + rep.TurnID})
+		return
+	}
 	if ok {
 		delete(r.taken, rep.TurnID)
 	}
@@ -295,6 +310,22 @@ func (r *Responder) Status() map[string]any {
 	return out
 }
 
+func randHex(n int) string {
+	b := make([]byte, n)
+	if _, err := rand.Read(b); err != nil {
+		panic("crypto/rand: " + err.Error())
+	}
+	return hex.EncodeToString(b)
+}
+
+// Handler serves the two endpoints; mount it on the dedicated external listener only.
+func (r *Responder) Handler() http.Handler {
+	mux := http.NewServeMux()
+	mux.HandleFunc("GET /external/next", r.Next)
+	mux.HandleFunc("POST /external/reply", r.ReplyHandler)
+	return mux
+}
+
 func writeJSON(w http.ResponseWriter, code int, v any) {
 	w.Header().Set("Content-Type", "application/json")
 	w.WriteHeader(code)
diff --git a/ai/internal/external/responder_test.go b/ai/internal/external/responder_test.go
index 06cd2d2..d9c9cfb 100644
--- a/ai/internal/external/responder_test.go
+++ b/ai/internal/external/responder_test.go
@@ -18,10 +18,7 @@ import (
 // the forced-submit flag; the reply's arguments may be a JSON string or object.
 func TestTurnsOverHTTP(t *testing.T) {
 	r := New(5 * time.Second)
-	mux := http.NewServeMux()
-	mux.HandleFunc("GET /external/next", r.Next)
-	mux.HandleFunc("POST /external/reply", r.ReplyHandler)
-	srv := httptest.NewServer(mux)
+	srv := httptest.NewServer(r.Handler())
 	defer srv.Close()
 	tools := []openai.Tool{{Type: "function", Strict: true, Name: "get_probe_status", Parameters: map[string]any{"type": "object"}}}
 	type res struct {
@@ -35,14 +32,18 @@ func TestTurnsOverHTTP(t *testing.T) {
 		ch <- res{resp, err}
 	}()
 	turn := poll(t, srv.URL)
-	if turn.TurnID != "ext_1" || turn.Session != "" || turn.Turn != 1 || turn.System != "SYS" || len(turn.Input) != 1 || turn.Input[0].Content != "alert" || len(turn.Tools) != 1 || turn.ForceSubmit || turn.MaxOutputTokens != 4000 {
+	if !strings.HasPrefix(turn.TurnID, "ext_") || len(turn.TurnID) != 20 || len(turn.ReplyToken) != 32 || turn.Session != "" || turn.Turn != 1 || turn.System != "SYS" || len(turn.Input) != 1 || turn.Input[0].Content != "alert" || len(turn.Tools) != 1 || turn.ForceSubmit || turn.MaxOutputTokens != 4000 {
 		t.Fatalf("turn 1: %+v", turn)
 	}
 	// second poll while nothing is queued -> 204 within the wait
 	if code := get(t, srv.URL+"/external/next?wait=0"); code != 204 {
 		t.Fatalf("empty queue must answer 204, got %d", code)
 	}
-	code, body := post(t, srv.URL+"/external/reply", `{"turn_id":"ext_1","output":[{"type":"message","text":"checking"},{"type":"function_call","name":"get_probe_status","arguments":"{}"}]}`)
+	// a reply without the turn's token is refused (the turn stays open)
+	if code, _ := post(t, srv.URL+"/external/reply", `{"turn_id":"`+turn.TurnID+`","reply_token":"wrong","output":[{"type":"message","text":"x"}]}`); code != 403 {
+		t.Fatalf("wrong token must be 403, got %d", code)
+	}
+	code, body := post(t, srv.URL+"/external/reply", `{"turn_id":"`+turn.TurnID+`","reply_token":"`+turn.ReplyToken+`","output":[{"type":"message","text":"checking"},{"type":"function_call","name":"get_probe_status","arguments":"{}"}]}`)
 	if code != 200 {
 		t.Fatalf("reply 1: %d %s", code, body)
 	}
@@ -51,7 +52,7 @@ func TestTurnsOverHTTP(t *testing.T) {
 		t.Fatal(r1.err)
 	}
 	fcs := r1.resp.FunctionCalls()
-	if r1.resp.ID != "ext_1" || len(fcs) != 1 || fcs[0].Name != "get_probe_status" || fcs[0].Arguments != "{}" || fcs[0].CallID == "" || r1.resp.Usage.TotalTokens != 0 {
+	if r1.resp.ID != turn.TurnID || len(fcs) != 1 || fcs[0].Name != "get_probe_status" || fcs[0].Arguments != "{}" || fcs[0].CallID == "" || r1.resp.Usage.TotalTokens != 0 {
 		t.Fatalf("response 1: %+v", r1.resp)
 	}
 	if !strings.Contains(r1.resp.Text(), "checking") {
@@ -59,20 +60,20 @@ func TestTurnsOverHTTP(t *testing.T) {
 	}
 
 	go func() {
-		resp, err := r.CreateResponse(context.Background(), openai.Request{Model: "external-claude-mythos-5-1", Tools: tools, PreviousResponseID: "ext_1",
+		resp, err := r.CreateResponse(context.Background(), openai.Request{Model: "external-claude-mythos-5-1", Tools: tools, PreviousResponseID: turn.TurnID,
 			ToolChoice: map[string]any{"type": "function", "name": "submit_diagnosis"},
 			Input:      []openai.Item{{Type: "function_call_output", CallID: fcs[0].CallID, Output: `{"alb":1}`}}})
 		ch <- res{resp, err}
 	}()
 	turn2 := poll(t, srv.URL)
-	if turn2.TurnID != "ext_2" || turn2.Session != "ext_1" || turn2.Turn != 2 || turn2.System != "SYS" || !turn2.ForceSubmit {
+	if turn2.TurnID == turn.TurnID || turn2.Session != turn.TurnID || turn2.Turn != 2 || turn2.System != "SYS" || !turn2.ForceSubmit {
 		t.Fatalf("turn 2 header: %+v", turn2)
 	}
 	// history: user alert / assistant text / assistant function_call / function_call_output
 	if len(turn2.Input) != 4 || turn2.Input[2].Type != "function_call" || turn2.Input[2].CallID != fcs[0].CallID || turn2.Input[3].Type != "function_call_output" || turn2.Input[3].Output != `{"alb":1}` {
 		t.Fatalf("turn 2 history: %s", mustJSON(turn2.Input))
 	}
-	code, body = post(t, srv.URL+"/external/reply", `{"turn_id":"ext_2","output":[{"type":"function_call","name":"submit_diagnosis","arguments":{"root_cause_label":"x","confidence":0.9}}]}`)
+	code, body = post(t, srv.URL+"/external/reply", `{"turn_id":"`+turn2.TurnID+`","reply_token":"`+turn2.ReplyToken+`","output":[{"type":"function_call","name":"submit_diagnosis","arguments":{"root_cause_label":"x","confidence":0.9}}]}`)
 	if code != 200 {
 		t.Fatalf("reply 2: %d %s", code, body)
 	}
@@ -84,7 +85,7 @@ func TestTurnsOverHTTP(t *testing.T) {
 		t.Fatalf("response 2: %+v", r2.resp.Output)
 	}
 	// an answered turn cannot be answered twice; malformed replies are refused before touching the queue
-	if code, _ := post(t, srv.URL+"/external/reply", `{"turn_id":"ext_2","output":[{"type":"message","text":"again"}]}`); code != 404 {
+	if code, _ := post(t, srv.URL+"/external/reply", `{"turn_id":"`+turn2.TurnID+`","reply_token":"`+turn2.ReplyToken+`","output":[{"type":"message","text":"again"}]}`); code != 404 {
 		t.Fatalf("second reply must be 404, got %d", code)
 	}
 	if code, _ := post(t, srv.URL+"/external/reply", `{"turn_id":"ext_9","output":[{"type":"function_call","name":"x","arguments":"[1]"}]}`); code != 400 {
@@ -118,10 +119,7 @@ func TestDeadlinesAndErrors(t *testing.T) {
 	}
 	// error reply
 	r2 := New(5 * time.Second)
-	mux := http.NewServeMux()
-	mux.HandleFunc("GET /external/next", r2.Next)
-	mux.HandleFunc("POST /external/reply", r2.ReplyHandler)
-	srv := httptest.NewServer(mux)
+	srv := httptest.NewServer(r2.Handler())
 	defer srv.Close()
 	ch := make(chan error, 1)
 	go func() {
@@ -129,7 +127,7 @@ func TestDeadlinesAndErrors(t *testing.T) {
 		ch <- err
 	}()
 	turn := poll(t, srv.URL)
-	post(t, srv.URL+"/external/reply", `{"turn_id":"`+turn.TurnID+`","error":"gave up"}`)
+	post(t, srv.URL+"/external/reply", `{"turn_id":"`+turn.TurnID+`","reply_token":"`+turn.ReplyToken+`","error":"gave up"}`)
 	if err := <-ch; err == nil || !strings.Contains(err.Error(), "gave up") {
 		t.Fatalf("expected client error, got %v", err)
 	}