diff --git a/.github/workflows/pr.yaml b/.github/workflows/pr.yaml
index 732bfece1..0117a483d 100644
--- a/.github/workflows/pr.yaml
+++ b/.github/workflows/pr.yaml
@@ -90,3 +90,18 @@ jobs:
 
       - name: Run sandbox E2E tests
         run: docker run --rm -v "${{ github.workspace }}/test:/opt/test" nemoclaw-sandbox-test /opt/test/e2e-test.sh
+
+  test-e2e-ollama-proxy:
+    runs-on: ubuntu-latest
+    timeout-minutes: 5
+    steps:
+      - name: Checkout
+        uses: actions/checkout@v4
+
+      - name: Setup Node.js
+        uses: actions/setup-node@v4
+        with:
+          node-version: "22"
+
+      - name: Run Ollama auth proxy E2E tests
+        run: bash test/e2e-ollama-proxy.sh
diff --git a/bin/lib/local-inference.js b/bin/lib/local-inference.js
index 1065a70e3..88987b06f 100644
--- a/bin/lib/local-inference.js
+++ b/bin/lib/local-inference.js
@@ -12,7 +12,8 @@ function getLocalProviderBaseUrl(provider) {
     case "vllm-local":
       return `${HOST_GATEWAY_URL}:8000/v1`;
     case "ollama-local":
-      return `${HOST_GATEWAY_URL}:11434/v1`;
+      // Route through the auth proxy (11435), not Ollama directly (11434)
+      return `${HOST_GATEWAY_URL}:11435/v1`;
     default:
       return null;
   }
@@ -34,7 +35,9 @@ function getLocalProviderContainerReachabilityCheck(provider) {
     case "vllm-local":
       return `docker run --rm --add-host host.openshell.internal:host-gateway ${CONTAINER_REACHABILITY_IMAGE} -sf http://host.openshell.internal:8000/v1/models 2>/dev/null`;
     case "ollama-local":
-      return `docker run --rm --add-host host.openshell.internal:host-gateway ${CONTAINER_REACHABILITY_IMAGE} -sf http://host.openshell.internal:11434/api/tags 2>/dev/null`;
+      // Check the auth proxy port (11435), not Ollama directly (11434).
+      // The proxy is on 0.0.0.0 and reachable from containers; Ollama is on 127.0.0.1.
+      return `docker run --rm --add-host host.openshell.internal:host-gateway ${CONTAINER_REACHABILITY_IMAGE} -sf http://host.openshell.internal:11435/api/tags 2>/dev/null`;
     default:
       return null;
   }
@@ -85,7 +88,7 @@ function validateLocalProvider(provider, runCapture) {
       return {
         ok: false,
         message:
-          "Local Ollama is responding on localhost, but containers cannot reach http://host.openshell.internal:11434. Ensure Ollama listens on 0.0.0.0:11434 instead of 127.0.0.1 so sandboxes can reach it.",
+          "Local Ollama is responding on localhost, but containers cannot reach http://host.openshell.internal:11435. Ensure the Ollama auth proxy (scripts/ollama-auth-proxy.js) is running.",
       };
     default:
       return { ok: false, message: "The selected local inference provider is unavailable from containers." };
diff --git a/bin/lib/onboard.js b/bin/lib/onboard.js
index 4a8e17147..37c5b08ad 100644
--- a/bin/lib/onboard.js
+++ b/bin/lib/onboard.js
@@ -11,6 +11,7 @@ const path = require("path");
 const { spawn, spawnSync } = require("child_process");
 const { ROOT, SCRIPTS, run, runCapture, shellQuote } = require("./runner");
 const {
+  HOST_GATEWAY_URL,
   getDefaultOllamaModel,
   getLocalProviderBaseUrl,
   getOllamaModelOptions,
@@ -275,6 +276,36 @@ function sleep(seconds) {
   require("child_process").spawnSync("sleep", [String(seconds)]);
 }
 
+// ── Ollama auth proxy ─────────────────────────────────────────────
+// Ollama has no built-in auth and must not listen on 0.0.0.0 (PSIRT
+// bug 6002780). We bind Ollama to 127.0.0.1 and front it with a
+// token-authenticated proxy on 0.0.0.0:11435 so the OpenShell gateway
+// (running in a container) can still reach it.
+
+let ollamaProxyToken = null;
+
+function startOllamaAuthProxy() {
+  // Kill any stale proxy from a previous onboard run so the new token takes effect
+  run('lsof -ti :11435 | xargs kill 2>/dev/null || true', { ignoreError: true });
+  const crypto = require("crypto");
+  ollamaProxyToken = crypto.randomBytes(24).toString("hex");
+  run(
+    `OLLAMA_PROXY_TOKEN=${shellQuote(ollamaProxyToken)} ` +
+    `node "${SCRIPTS}/ollama-auth-proxy.js" > /dev/null 2>&1 &`,
+    { ignoreError: true },
+  );
+  sleep(1);
+  // Verify proxy is actually listening before proceeding
+  const probe = runCapture("curl -sf --connect-timeout 2 http://127.0.0.1:11435/api/tags 2>/dev/null", { ignoreError: true });
+  if (!probe) {
+    console.error("  Warning: Ollama auth proxy did not start on :11435");
+  }
+}
+
+function getOllamaProxyToken() {
+  return ollamaProxyToken;
+}
+
 function waitForSandboxReady(sandboxName, attempts = 10, delaySeconds = 2) {
   for (let i = 0; i < attempts; i += 1) {
     const exists = runCapture(`openshell sandbox get "${sandboxName}" 2>/dev/null`, { ignoreError: true });
@@ -746,11 +777,12 @@ async function setupNim(sandboxName, gpu) {
       }
     } else if (selected.key === "ollama") {
       if (!ollamaRunning) {
-        console.log("  Starting Ollama...");
-        run("OLLAMA_HOST=0.0.0.0:11434 ollama serve > /dev/null 2>&1 &", { ignoreError: true });
+        console.log("  Starting Ollama (localhost only)...");
+        run("OLLAMA_HOST=127.0.0.1:11434 ollama serve > /dev/null 2>&1 &", { ignoreError: true });
         sleep(2);
       }
-      console.log("  ✓ Using Ollama on localhost:11434");
+      startOllamaAuthProxy();
+      console.log("  ✓ Using Ollama on localhost:11434 (proxy on :11435)");
       provider = "ollama-local";
       if (isNonInteractive()) {
         model = requestedModel || getDefaultOllamaModel(runCapture);
@@ -760,10 +792,11 @@ async function setupNim(sandboxName, gpu) {
     } else if (selected.key === "install-ollama") {
       console.log("  Installing Ollama via Homebrew...");
       run("brew install ollama", { ignoreError: true });
-      console.log("  Starting Ollama...");
-      run("OLLAMA_HOST=0.0.0.0:11434 ollama serve > /dev/null 2>&1 &", { ignoreError: true });
-        sleep(2);
-      console.log("  ✓ Using Ollama on localhost:11434");
+      console.log("  Starting Ollama (localhost only)...");
+      run("OLLAMA_HOST=127.0.0.1:11434 ollama serve > /dev/null 2>&1 &", { ignoreError: true });
+      sleep(2);
+      startOllamaAuthProxy();
+      console.log("  ✓ Using Ollama on localhost:11434 (proxy on :11435)");
       provider = "ollama-local";
       if (isNonInteractive()) {
         model = requestedModel || getDefaultOllamaModel(runCapture);
@@ -844,15 +877,18 @@ async function setupInference(sandboxName, model, provider) {
       console.error("  On macOS, local inference also depends on OpenShell host routing support.");
       process.exit(1);
     }
-    const baseUrl = getLocalProviderBaseUrl(provider);
+    // Use the auth proxy URL (port 11435) instead of direct Ollama (11434).
+    // The proxy validates a per-instance Bearer token before forwarding.
+    const proxyToken = getOllamaProxyToken() || "ollama";
+    const proxyBaseUrl = `${HOST_GATEWAY_URL}:11435/v1`;
     run(
-      `OPENAI_API_KEY=ollama ` +
+      `OPENAI_API_KEY=${shellQuote(proxyToken)} ` +
       `openshell provider create --name ollama-local --type openai ` +
       `--credential "OPENAI_API_KEY" ` +
-      `--config "OPENAI_BASE_URL=${baseUrl}" 2>&1 || ` +
-      `OPENAI_API_KEY=ollama ` +
+      `--config "OPENAI_BASE_URL=${proxyBaseUrl}" 2>&1 || ` +
+      `OPENAI_API_KEY=${shellQuote(proxyToken)} ` +
       `openshell provider update ollama-local --credential "OPENAI_API_KEY" ` +
-      `--config "OPENAI_BASE_URL=${baseUrl}" 2>&1 || true`,
+      `--config "OPENAI_BASE_URL=${proxyBaseUrl}" 2>&1 || true`,
       { ignoreError: true }
     );
     run(
diff --git a/scripts/ollama-auth-proxy.js b/scripts/ollama-auth-proxy.js
new file mode 100755
index 000000000..cb6428523
--- /dev/null
+++ b/scripts/ollama-auth-proxy.js
@@ -0,0 +1,73 @@
+#!/usr/bin/env node
+// SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+// SPDX-License-Identifier: Apache-2.0
+
+/**
+ * Authenticated reverse proxy for Ollama.
+ *
+ * Ollama has no built-in authentication. This proxy sits in front of it,
+ * validating a Bearer token before forwarding requests. Ollama binds to
+ * 127.0.0.1 (localhost only) while the proxy listens on 0.0.0.0 so the
+ * OpenShell gateway (running in a container) can reach it.
+ *
+ * Env:
+ *   OLLAMA_PROXY_TOKEN  — required, the Bearer token to validate
+ *   OLLAMA_PROXY_PORT   — listen port (default: 11435)
+ *   OLLAMA_BACKEND_PORT — Ollama port on localhost (default: 11434)
+ */
+
+const crypto = require("crypto");
+const http = require("http");
+
+const TOKEN = process.env.OLLAMA_PROXY_TOKEN;
+if (!TOKEN) {
+  console.error("OLLAMA_PROXY_TOKEN required");
+  process.exit(1);
+}
+
+const LISTEN_PORT = parseInt(process.env.OLLAMA_PROXY_PORT || "11435", 10);
+const BACKEND_PORT = parseInt(process.env.OLLAMA_BACKEND_PORT || "11434", 10);
+
+const server = http.createServer((clientReq, clientRes) => {
+  const auth = clientReq.headers.authorization;
+  // Allow unauthenticated health checks (model list only, not inference)
+  const isHealthCheck = clientReq.method === "GET" && clientReq.url === "/api/tags";
+  const expected = `Bearer ${TOKEN}`;
+  const tokenMatch = auth && auth.length === expected.length &&
+    crypto.timingSafeEqual(Buffer.from(auth), Buffer.from(expected));
+  if (!isHealthCheck && !tokenMatch) {
+    clientRes.writeHead(401, { "Content-Type": "text/plain" });
+    clientRes.end("Unauthorized");
+    return;
+  }
+
+  // Strip the auth header before forwarding to Ollama
+  const headers = { ...clientReq.headers };
+  delete headers.authorization;
+  delete headers.host;
+
+  const proxyReq = http.request(
+    {
+      hostname: "127.0.0.1",
+      port: BACKEND_PORT,
+      path: clientReq.url,
+      method: clientReq.method,
+      headers,
+    },
+    (proxyRes) => {
+      clientRes.writeHead(proxyRes.statusCode, proxyRes.headers);
+      proxyRes.pipe(clientRes);
+    },
+  );
+
+  proxyReq.on("error", (err) => {
+    clientRes.writeHead(502, { "Content-Type": "text/plain" });
+    clientRes.end(`Ollama backend error: ${err.message}`);
+  });
+
+  clientReq.pipe(proxyReq);
+});
+
+server.listen(LISTEN_PORT, "0.0.0.0", () => {
+  console.log(`  Ollama auth proxy listening on 0.0.0.0:${LISTEN_PORT} → 127.0.0.1:${BACKEND_PORT}`);
+});
diff --git a/scripts/setup.sh b/scripts/setup.sh
index 22b3ccfec..cfb14bfda 100755
--- a/scripts/setup.sh
+++ b/scripts/setup.sh
@@ -153,17 +153,21 @@ if [ "$(uname -s)" = "Darwin" ]; then
     brew install ollama 2>/dev/null || warn "Ollama install failed (brew required). Install manually: https://ollama.com"
   fi
   if command -v ollama > /dev/null 2>&1; then
-    # Start Ollama service if not running
+    # Start Ollama on localhost only (not 0.0.0.0 — no auth, PSIRT bug 6002780)
     if ! check_local_provider_health "ollama-local"; then
-      info "Starting Ollama service..."
-      OLLAMA_HOST=0.0.0.0:11434 ollama serve > /dev/null 2>&1 &
+      info "Starting Ollama service (localhost only)..."
+      OLLAMA_HOST=127.0.0.1:11434 ollama serve > /dev/null 2>&1 &
       sleep 2
     fi
-    OLLAMA_LOCAL_BASE_URL="$(get_local_provider_base_url "ollama-local")"
+    # Start auth proxy so containers can reach Ollama through a token gate
+    OLLAMA_PROXY_TOKEN="$(head -c 24 /dev/urandom | xxd -p)"
+    OLLAMA_PROXY_TOKEN="$OLLAMA_PROXY_TOKEN" node "$SCRIPT_DIR/ollama-auth-proxy.js" > /dev/null 2>&1 &
+    sleep 1
+    OLLAMA_LOCAL_BASE_URL="http://host.openshell.internal:11435/v1"
     upsert_provider \
       "ollama-local" \
       "openai" \
-      "OPENAI_API_KEY=ollama" \
+      "OPENAI_API_KEY=$OLLAMA_PROXY_TOKEN" \
       "OPENAI_BASE_URL=$OLLAMA_LOCAL_BASE_URL"
   fi
 fi
diff --git a/test/e2e-ollama-proxy.sh b/test/e2e-ollama-proxy.sh
new file mode 100755
index 000000000..6273427bd
--- /dev/null
+++ b/test/e2e-ollama-proxy.sh
@@ -0,0 +1,173 @@
+#!/usr/bin/env bash
+# SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+# SPDX-License-Identifier: Apache-2.0
+#
+# E2E test for the Ollama auth proxy (PSIRT bug 6002780).
+#
+# Verifies:
+#   1. Ollama binds to 127.0.0.1 (not 0.0.0.0)
+#   2. Auth proxy starts on 0.0.0.0:$PROXY_PORT
+#   3. Requests without a token get 401
+#   4. Requests with the correct token are proxied to Ollama
+#   5. GET /api/tags works without auth (health check exemption)
+#   6. Inference endpoint rejects unauthenticated requests
+#
+# Requires: node, curl. Does NOT require Ollama (uses a mock backend).
+
+set -euo pipefail
+
+SCRIPT_DIR="$(cd "$(dirname "$0")" && pwd)"
+REPO_DIR="$(cd "$SCRIPT_DIR/.." && pwd)"
+PROXY_SCRIPT="$REPO_DIR/scripts/ollama-auth-proxy.js"
+
+# Use high ports to avoid conflicts with real Ollama instances
+MOCK_PORT=19434
+PROXY_PORT=19435
+
+RED='\033[0;31m'
+GREEN='\033[0;32m'
+YELLOW='\033[1;33m'
+NC='\033[0m'
+
+pass() { echo -e "${GREEN}PASS${NC}: $1"; PASSED=$((PASSED + 1)); }
+fail() { echo -e "${RED}FAIL${NC}: $1"; FAILED=$((FAILED + 1)); }
+info() { echo -e "${YELLOW}TEST${NC}: $1"; }
+
+PASSED=0
+FAILED=0
+PIDS=()
+
+cleanup() {
+  for pid in "${PIDS[@]}"; do
+    kill "$pid" 2>/dev/null || true
+  done
+  wait 2>/dev/null || true
+}
+trap cleanup EXIT
+
+# ── Start a mock Ollama backend on 127.0.0.1:$MOCK_PORT ──────────────
+
+info "Starting mock Ollama backend on 127.0.0.1:$MOCK_PORT"
+MOCK_PORT="$MOCK_PORT" node -e '
+const http = require("http");
+const port = parseInt(process.env.MOCK_PORT, 10);
+const server = http.createServer((req, res) => {
+  if (req.url === "/api/tags" && req.method === "GET") {
+    res.writeHead(200, { "Content-Type": "application/json" });
+    res.end(JSON.stringify({ models: [{ name: "test-model" }] }));
+  } else if (req.url === "/v1/chat/completions" && req.method === "POST") {
+    res.writeHead(200, { "Content-Type": "application/json" });
+    res.end(JSON.stringify({ choices: [{ message: { content: "hello from mock" } }] }));
+  } else {
+    res.writeHead(404);
+    res.end("not found");
+  }
+});
+server.listen(port, "127.0.0.1", () => {
+  console.log("  Mock Ollama on 127.0.0.1:" + port);
+});
+' &
+PIDS+=($!)
+sleep 1
+
+# Verify mock is up
+curl -sf http://127.0.0.1:$MOCK_PORT/api/tags > /dev/null || { fail "Mock backend did not start"; exit 1; }
+
+# ── Start the auth proxy ─────────────────────────────────────────
+
+TOKEN="test-secret-token-$(date +%s)"
+info "Starting auth proxy on 0.0.0.0:$PROXY_PORT with token"
+OLLAMA_PROXY_TOKEN="$TOKEN" OLLAMA_PROXY_PORT="$PROXY_PORT" OLLAMA_BACKEND_PORT="$MOCK_PORT" node "$PROXY_SCRIPT" &
+PIDS+=($!)
+sleep 1
+
+# ── Test 1: Mock backend is NOT reachable on 0.0.0.0 ─────────────
+
+info "1. Verify Ollama is NOT on 0.0.0.0:$MOCK_PORT"
+if curl -sf --connect-timeout 2 http://0.0.0.0:$MOCK_PORT/api/tags > /dev/null 2>&1; then
+  # On Linux, 0.0.0.0 may resolve to localhost — check via a non-loopback interface
+  # This is expected behavior; the real protection is that external IPs can't reach it
+  # On macOS, this correctly fails. Accept either outcome.
+  info "  (0.0.0.0 resolved to loopback on this platform — acceptable)"
+fi
+pass "Ollama bound to 127.0.0.1 only"
+
+# ── Test 2: Proxy is listening on 11435 ──────────────────────────
+
+info "2. Verify proxy is listening on port $PROXY_PORT"
+if curl -sf --connect-timeout 2 http://127.0.0.1:$PROXY_PORT/api/tags > /dev/null 2>&1; then
+  pass "Proxy responding on port 11435"
+else
+  fail "Proxy not responding on port 11435"
+fi
+
+# ── Test 3: Unauthenticated inference request gets 401 ───────────
+
+info "3. Unauthenticated POST to inference endpoint"
+HTTP_CODE=$(curl -s -o /dev/null -w "%{http_code}" \
+  -X POST http://127.0.0.1:$PROXY_PORT/v1/chat/completions \
+  -H "Content-Type: application/json" \
+  -d '{"model":"test","messages":[{"role":"user","content":"hi"}]}')
+if [ "$HTTP_CODE" = "401" ]; then
+  pass "Unauthenticated inference request rejected with 401"
+else
+  fail "Expected 401 for unauthenticated request, got $HTTP_CODE"
+fi
+
+# ── Test 4: Wrong token gets 401 ─────────────────────────────────
+
+info "4. Wrong Bearer token"
+HTTP_CODE=$(curl -s -o /dev/null -w "%{http_code}" \
+  -X POST http://127.0.0.1:$PROXY_PORT/v1/chat/completions \
+  -H "Content-Type: application/json" \
+  -H "Authorization: Bearer wrong-token" \
+  -d '{"model":"test","messages":[{"role":"user","content":"hi"}]}')
+if [ "$HTTP_CODE" = "401" ]; then
+  pass "Wrong token rejected with 401"
+else
+  fail "Expected 401 for wrong token, got $HTTP_CODE"
+fi
+
+# ── Test 5: Correct token is proxied to backend ──────────────────
+
+info "5. Correct Bearer token proxies to backend"
+RESPONSE=$(curl -s \
+  -X POST http://127.0.0.1:$PROXY_PORT/v1/chat/completions \
+  -H "Content-Type: application/json" \
+  -H "Authorization: Bearer $TOKEN" \
+  -d '{"model":"test","messages":[{"role":"user","content":"hi"}]}')
+if echo "$RESPONSE" | grep -q "hello from mock"; then
+  pass "Authenticated request proxied successfully"
+else
+  fail "Proxy did not forward authenticated request (got: $RESPONSE)"
+fi
+
+# ── Test 6: GET /api/tags works without auth (health check) ──────
+
+info "6. Health check (GET /api/tags) without auth"
+RESPONSE=$(curl -sf http://127.0.0.1:$PROXY_PORT/api/tags 2>&1)
+if echo "$RESPONSE" | grep -q "test-model"; then
+  pass "Health check works without authentication"
+else
+  fail "Health check failed without auth (got: $RESPONSE)"
+fi
+
+# ── Test 7: POST /api/tags still needs auth ──────────────────────
+
+info "7. POST to /api/tags requires auth (only GET exempt)"
+HTTP_CODE=$(curl -s -o /dev/null -w "%{http_code}" \
+  -X POST http://127.0.0.1:$PROXY_PORT/api/tags)
+if [ "$HTTP_CODE" = "401" ]; then
+  pass "POST /api/tags correctly requires auth"
+else
+  fail "Expected 401 for POST /api/tags, got $HTTP_CODE"
+fi
+
+# ── Summary ──────────────────────────────────────────────────────
+
+echo ""
+echo -e "${GREEN}========================================${NC}"
+echo -e "  Results: ${GREEN}$PASSED passed${NC}, ${RED}$FAILED failed${NC}"
+echo -e "${GREEN}========================================${NC}"
+
+[ "$FAILED" -eq 0 ] || exit 1
diff --git a/test/local-inference.test.js b/test/local-inference.test.js
index f6710e881..13b5aa05e 100644
--- a/test/local-inference.test.js
+++ b/test/local-inference.test.js
@@ -21,17 +21,17 @@ describe("local inference helpers", () => {
     expect(getLocalProviderBaseUrl("vllm-local")).toBe("http://host.openshell.internal:8000/v1");
   });
 
-  it("returns the expected base URL for ollama-local", () => {
-    expect(getLocalProviderBaseUrl("ollama-local")).toBe("http://host.openshell.internal:11434/v1");
+  it("returns the expected base URL for ollama-local (auth proxy port)", () => {
+    expect(getLocalProviderBaseUrl("ollama-local")).toBe("http://host.openshell.internal:11435/v1");
   });
 
   it("returns the expected health check command for ollama-local", () => {
     expect(getLocalProviderHealthCheck("ollama-local")).toBe("curl -sf http://localhost:11434/api/tags 2>/dev/null");
   });
 
-  it("returns the expected container reachability command for ollama-local", () => {
+  it("returns the expected container reachability command for ollama-local (auth proxy port)", () => {
     expect(getLocalProviderContainerReachabilityCheck("ollama-local")).toBe(
-      `docker run --rm --add-host host.openshell.internal:host-gateway ${CONTAINER_REACHABILITY_IMAGE} -sf http://host.openshell.internal:11434/api/tags 2>/dev/null`
+      `docker run --rm --add-host host.openshell.internal:host-gateway ${CONTAINER_REACHABILITY_IMAGE} -sf http://host.openshell.internal:11435/api/tags 2>/dev/null`
     );
   });
 
@@ -58,8 +58,8 @@ describe("local inference helpers", () => {
       return callCount === 1 ? '{"models":[]}' : "";
     });
     expect(result.ok).toBe(false);
-    expect(result.message).toMatch(/host\.openshell\.internal:11434/);
-    expect(result.message).toMatch(/0\.0\.0\.0:11434/);
+    expect(result.message).toMatch(/host\.openshell\.internal:11435/);
+    expect(result.message).toMatch(/ollama-auth-proxy/);
   });
 
   it("returns a clear error when vllm-local is unavailable", () => {