diff --git a/.env.example b/.env.example
index aa5a7e3c3..6056279ba 100644
--- a/.env.example
+++ b/.env.example
@@ -109,10 +109,14 @@ API_TOKEN_HMAC_KEY="REPLACE-WITH-OUTPUT-OF-openssl-rand-hex-32"
 # Users bring their own API keys via Settings. No server-side key needed by
 # default. The admin can configure a fallback key in the Admin Panel.
 
-# Local-AI SSRF opt-in (optional). The local AI provider rejects private/
-# internal hosts by default. Set to an exact host allowlist (comma-separated,
-# e.g. "ollama.lan,10.0.0.5") or "true" for any private host — intended for a
-# self-hosted Ollama / LM Studio endpoint. See docs/integrations/ai-providers.md.
+# AI endpoints on your own network (optional). An AI base URL on a private
+# address is refused unless its exact origin is listed here, comma-separated
+# (e.g. "http://ollama.lan:11434,http://10.0.0.5:4000") — for a self-hosted
+# Ollama, LM Studio or LiteLLM gateway. Link-local, metadata and unspecified
+# addresses stay refused even when listed. See docs/integrations/ai-providers.md.
+# AI_PRIVATE_ORIGINS=""
+# Older form, kept for existing installs: an exact host list still works
+# (e.g. "ollama.lan,10.0.0.5"). "true" no longer opens every private host.
 # ALLOW_LOCAL_AI_PRIVATE_HOSTS=""
 
 # Daily-briefing output-token ceiling (optional). Default 2500, clamped to
diff --git a/.env.production.example b/.env.production.example
index ead149065..605ccd7e3 100644
--- a/.env.production.example
+++ b/.env.production.example
@@ -267,10 +267,14 @@ API_TOKEN_HMAC_KEY="<openssl rand -hex 32>"
 # docs/ops/tls-cert-pin.md for extraction + the re-pin runbook.
 # TLS_LEAF_SPKI_PINS=""
 
-# Local-AI SSRF opt-in (optional). The local AI provider rejects private/
-# internal hosts by default. Set to an exact host allowlist (comma-separated,
-# e.g. "ollama.lan,10.0.0.5") or "true" for any private host — intended for a
-# self-hosted Ollama / LM Studio endpoint. See docs/integrations/ai-providers.md.
+# AI endpoints on your own network (optional). An AI base URL on a private
+# address is refused unless its exact origin is listed here, comma-separated
+# (e.g. "http://ollama.lan:11434,http://10.0.0.5:4000") — for a self-hosted
+# Ollama, LM Studio or LiteLLM gateway. Link-local, metadata and unspecified
+# addresses stay refused even when listed. See docs/integrations/ai-providers.md.
+# AI_PRIVATE_ORIGINS=""
+# Older form, kept for existing installs: an exact host list still works
+# (e.g. "ollama.lan,10.0.0.5"). "true" no longer opens every private host.
 # ALLOW_LOCAL_AI_PRIVATE_HOSTS=""
 
 # Daily-briefing output-token ceiling (optional). Default 2500, clamped to
diff --git a/docker-compose.yml b/docker-compose.yml
index a24b90ba0..797bd2d48 100644
--- a/docker-compose.yml
+++ b/docker-compose.yml
@@ -253,12 +253,16 @@ x-healthlog-env: &healthlog_env
   # container (the `environment:` block is a whitelist). See
   # docs/ops/tls-cert-pin.md.
   TLS_LEAF_SPKI_PINS: "${TLS_LEAF_SPKI_PINS:-}"
-  # Local-AI SSRF opt-in (optional). The local AI provider rejects
-  # private/internal hosts by default; set this to an exact host allowlist
-  # (e.g. "ollama.lan,10.0.0.5") or "true" for any private host, intended
-  # for a self-hosted Ollama / LM Studio. Listed here so an operator value
-  # in .env reaches the container (the `environment:` block is a whitelist).
-  # See docs/integrations/ai-providers.md.
+  # AI endpoints on your own network (optional): a self-hosted Ollama,
+  # LM Studio or LiteLLM gateway, comma-separated exact origins (e.g.
+  # "http://ollama.lan:11434,http://10.0.0.5:4000"). Without it an AI base
+  # URL on a private address is refused. Link-local, metadata and
+  # unspecified addresses stay refused even when listed. See
+  # docs/integrations/ai-providers.md.
+  AI_PRIVATE_ORIGINS: "${AI_PRIVATE_ORIGINS:-}"
+  # Older form of the above, kept for existing installs: an exact host list
+  # (e.g. "ollama.lan,10.0.0.5") still works. Its "true" form no longer
+  # opens every private host; move to AI_PRIVATE_ORIGINS.
   ALLOW_LOCAL_AI_PRIVATE_HOSTS: "${ALLOW_LOCAL_AI_PRIVATE_HOSTS:-}"
   # Extra hosts the admin-configured AI base URL may point at, on top of the
   # built-in public providers. Comma-separated bare hostnames. Set it when
diff --git a/docs/integrations/ai-providers.md b/docs/integrations/ai-providers.md
index e4574e949..1ba9659a0 100644
--- a/docs/integrations/ai-providers.md
+++ b/docs/integrations/ai-providers.md
@@ -140,16 +140,28 @@ Base URL `http://<vllm-host>:8000/v1`, API key `your-shared-secret`,
 model `Qwen/Qwen2.5-14B-Instruct`. The local-client encrypts the
 API key at rest like every other provider credential.
 
-**SSRF guard.** Local endpoints route through the same validation
-as every other outbound URL HealthLog hits. The guard rejects
-private-IP-range targets by default. `ALLOW_LOCAL_AI_PRIVATE_HOSTS`
-opens the escape hatch and accepts two forms: a comma-separated host
-allowlist (e.g. `ALLOW_LOCAL_AI_PRIVATE_HOSTS=ollama.lan,10.0.0.5`),
-which permits only those exact hostnames, or the legacy
-`ALLOW_LOCAL_AI_PRIVATE_HOSTS=true`, which permits any private host —
-including the cloud-metadata endpoint, so prefer the host list. Leave
-it unset on public-facing instances unless you specifically run an
-internal endpoint.
+**Private endpoints.** Any user of the instance can type a base URL,
+so HealthLog refuses one on a private address unless the operator has
+listed it. List the endpoint's exact origin in `AI_PRIVATE_ORIGINS`,
+comma-separated, the same way `NOTIFICATION_PRIVATE_ORIGINS` works for
+notification relays:
+
+```bash
+AI_PRIVATE_ORIGINS="http://ollama.lan:11434,http://10.0.0.5:4000"
+```
+
+An origin is `scheme://host[:port]` and nothing else: no path, no
+wildcard, no range. A listed endpoint is called through a pinned
+connection that refuses to follow redirects, and the cloud-metadata
+address, link-local addresses and the unspecified address stay
+unreachable even when listed or when a listed name resolves to one.
+
+`ALLOW_LOCAL_AI_PRIVATE_HOSTS`, the older setting, still works as a
+host list (`ollama.lan,10.0.0.5`, any port), with the same pinning.
+Its `true` form used to open every private host on the network to
+every user and no longer grants anything since v1.39.3; the boot
+summary names it when it is still set. Replace it with the one origin
+your endpoint listens on.
 
 **Model sizing.** Roughly: 7-8B for 8 GB GPU / Apple Silicon base
 (briefings fine, deeper Coach reasoning gets terse), 14-24B for
@@ -186,9 +198,9 @@ therefore never be sent to a host you typed into a different form,
 and that property is structural rather than a comment — the OPENAI
 arm has no code path that reads the gateway's base URL.
 
-**SSRF guard.** The gateway's base URL passes the same floor as the
-Local provider's: a public host always, a private one only when the
-operator allowlisted it via `ALLOW_LOCAL_AI_PRIVATE_HOSTS` (documented
+**Private endpoints.** The gateway's base URL passes the same floor
+as the Local provider's: a public host always, a private one only when
+the operator listed its origin in `AI_PRIVATE_ORIGINS` (documented
 under Local endpoints above). LAN gateways are the normal case for LiteLLM and vLLM, so
 the escape hatch matters here as much as it does for Ollama.
 
@@ -439,12 +451,12 @@ upstream error.
 - **"No AI provider configured."** Either set up at least one BYOK
   provider in `/settings/ai`, or configure an admin-shared OpenAI
   key in the admin panel.
-- **Local endpoint rejected as "internal/private host".** Set
-  `ALLOW_LOCAL_AI_PRIVATE_HOSTS` to the endpoint's exact hostname
-  (e.g. `ollama.lan`; comma-separate several) in the container
-  environment, or to `true` to allow any private host. Prefer the host
-  list; leave it unset on internet-facing instances unless you
-  specifically want to allow private-range targets.
+- **Local endpoint rejected as "internal/private host".** Add the
+  endpoint's exact origin to `AI_PRIVATE_ORIGINS` in the container
+  environment (e.g. `http://ollama.lan:11434`; comma-separate several)
+  and restart. If it stopped working after an update and you had
+  `ALLOW_LOCAL_AI_PRIVATE_HOSTS=true`, that form no longer grants
+  anything; list the origin instead.
 - **Codex disconnects after a few weeks of inactivity.** ChatGPT
   refresh tokens lapse after extended idle periods. Reconnect via
   the **Connect ChatGPT** button — HealthLog re-runs the device-code
diff --git a/scripts/env-manifest.json b/scripts/env-manifest.json
index 11fcfadb5..f0b141d6e 100644
--- a/scripts/env-manifest.json
+++ b/scripts/env-manifest.json
@@ -122,9 +122,13 @@
       "description": "Optional runtime policy and output tuning for user-configured AI providers. Each variable stands alone and has a safe default when unset.",
       "required": false,
       "variables": [
+        {
+          "name": "AI_PRIVATE_ORIGINS",
+          "purpose": "AI endpoints on the operator's own network (Ollama, LM Studio, LiteLLM). Comma-separated exact origins (scheme://host[:port]), e.g. \"http://ollama.lan:11434\". Unset refuses every private AI base URL. Link-local, metadata and unspecified addresses stay refused even when listed; a malformed entry is logged once and grants nothing."
+        },
         {
           "name": "ALLOW_LOCAL_AI_PRIVATE_HOSTS",
-          "purpose": "Local-AI SSRF opt-in. Unset rejects private/internal hosts; set an exact comma-separated host allowlist or \"true\" to allow any private host."
+          "purpose": "Older form of AI_PRIVATE_ORIGINS, kept for existing installs: an exact comma-separated host list still grants those hosts. \"true\" grants nothing since v1.39.3 and is reported at boot."
         },
         {
           "name": "INSIGHTS_MAX_TOKENS",
diff --git a/src/__tests__/compose-env-whitelist-guard.test.ts b/src/__tests__/compose-env-whitelist-guard.test.ts
index 20c3ad68e..d85377847 100644
--- a/src/__tests__/compose-env-whitelist-guard.test.ts
+++ b/src/__tests__/compose-env-whitelist-guard.test.ts
@@ -23,10 +23,11 @@
  *
  * ## Why the matchers are what they are
  *
- * A grep for `process.env.X` alone is not the read set. Four modules resolve a
- * name through a helper and index `process.env` dynamically, so their reads
- * are invisible to the direct matcher — `ALLOW_LOCAL_AI_PRIVATE_HOSTS` and
- * every `OIDC_*` among them. Each is registered below with its own call-shape
+ * A grep for `process.env.X` alone is not the read set. Three modules resolve
+ * a name through a helper and index `process.env` dynamically, so their reads
+ * are invisible to the direct matcher — every `OIDC_*` among them. (The AI
+ * private-host grants were a fourth until v1.39.3, when they became direct
+ * `process.env` reads.) Each is registered below with its own call-shape
  * matcher AND its own non-zero assertion, so a refactor that renames the
  * helper fails this file loudly instead of quietly shrinking the read set.
  *
@@ -72,12 +73,6 @@ const INDIRECT_READERS: ReadonlyArray<{
     minimum: 1,
     why: "the retention windows are parsed through a local intEnv() helper",
   },
-  {
-    file: "lib/ai/local-host-allowlist.ts",
-    pattern: /ENV_VAR\s*=\s*["']([A-Z][A-Z0-9_]*)["']/g,
-    minimum: 1,
-    why: "the name is hoisted into an ENV_VAR constant, then indexed",
-  },
   {
     file: "lib/boot/readiness-summary.ts",
     pattern: /\benv\.([A-Z][A-Z0-9_]*)/g,
diff --git a/src/app/api/labs/ocr/extract/__tests__/route.test.ts b/src/app/api/labs/ocr/extract/__tests__/route.test.ts
index 66b874e13..4dcd20385 100644
--- a/src/app/api/labs/ocr/extract/__tests__/route.test.ts
+++ b/src/app/api/labs/ocr/extract/__tests__/route.test.ts
@@ -66,6 +66,7 @@ import { rasterizePdf } from "@/lib/documents/rasterize-pdf";
 import { reserveBudget, reconcileSpend } from "@/lib/ai/coach/budget";
 import { checkRateLimit, refundRateLimit } from "@/lib/rate-limit";
 import { OcrExtractError, runOcrExtraction } from "@/lib/labs/ocr-extract";
+import { annotate } from "@/lib/logging/context";
 
 const SESSION_OK = {
   session: { id: "sess-1", expiresAt: new Date(Date.now() + 3_600_000) },
@@ -164,6 +165,29 @@ describe("POST /api/labs/ocr/extract — text mode budget", () => {
     expect(res.status).toBe(502);
     expect(body.error).toContain("configured AI provider");
   });
+
+  it("never passes the upstream body on, to the caller or to the event", async () => {
+    const upstreamBody = "ami-id instance-id iam/security-credentials/admin";
+    vi.mocked(runOcrExtraction).mockRejectedValue(
+      Object.assign(new Error("Local AI request failed (403)"), {
+        httpStatus: 403,
+        model: "llama3",
+        bodyExcerpt: upstreamBody,
+      }),
+    );
+
+    const res = await POST(textReq());
+    expect(res.status).toBe(502);
+    expect(await res.text()).not.toContain("iam/security-credentials");
+    const events = JSON.stringify(vi.mocked(annotate).mock.calls);
+    expect(events).not.toContain("iam/security-credentials");
+    // The status and the model stay, so an operator can still tell why.
+    expect(annotate).toHaveBeenCalledWith(
+      expect.objectContaining({
+        meta: expect.objectContaining({ upstreamStatus: 403, model: "llama3" }),
+      }),
+    );
+  });
 });
 
 describe("POST /api/labs/ocr/extract — vision PDF rasterization", () => {
diff --git a/src/app/api/labs/ocr/extract/route.ts b/src/app/api/labs/ocr/extract/route.ts
index c6b1db112..16ca28416 100644
--- a/src/app/api/labs/ocr/extract/route.ts
+++ b/src/app/api/labs/ocr/extract/route.ts
@@ -72,11 +72,16 @@ export const dynamic = "force-dynamic";
 /** OCR'd text is bounded in the schema; cap the JSON body proportionally. */
 const TEXT_BODY_MAX_BYTES = 512 * 1024;
 
+/**
+ * The wide-event facts of a failed provider call: the status and the model,
+ * never the upstream body. The base URL can be one a user typed, and what an
+ * arbitrary endpoint answers has no business in the event stream (v1.39.3);
+ * the response to the caller carries neither.
+ */
 function providerFailureMeta(error: unknown, mode?: "text") {
   const err = error as {
     httpStatus?: unknown;
     model?: unknown;
-    bodyExcerpt?: unknown;
   };
   return {
     reason: "provider_error",
@@ -85,9 +90,6 @@ function providerFailureMeta(error: unknown, mode?: "text") {
       ? { upstreamStatus: err.httpStatus }
       : {}),
     ...(typeof err.model === "string" ? { model: err.model } : {}),
-    ...(typeof err.bodyExcerpt === "string"
-      ? { upstreamError: err.bodyExcerpt }
-      : {}),
   };
 }
 
diff --git a/src/app/api/user/ai-provider/route.ts b/src/app/api/user/ai-provider/route.ts
index 3aa4edf38..a10f708eb 100644
--- a/src/app/api/user/ai-provider/route.ts
+++ b/src/app/api/user/ai-provider/route.ts
@@ -197,14 +197,14 @@ export const PATCH = apiHandler(async (request: NextRequest) => {
       const trimmed = body.baseUrl.trim();
       // SSRF guard: by default reject private/internal hostnames so a
       // compromised user account cannot point the server at the cloud
-      // metadata endpoint or internal admin panels. v1.18.7 (SECURITY LOW) —
-      // ops opt in via `ALLOW_LOCAL_AI_PRIVATE_HOSTS`, now a host allowlist:
-      // `true` permits any private host (legacy), a comma-separated host list
-      // permits only those exact hostnames.
+      // metadata endpoint or internal admin panels. The operator grants a
+      // private endpoint by exact origin (`AI_PRIVATE_ORIGINS`, or the legacy
+      // host list); the retired `ALLOW_LOCAL_AI_PRIVATE_HOSTS=true` grants
+      // nothing since v1.39.3, and no grant can open metadata or link-local.
       const allowPrivate = isLocalAiHostAllowed(trimmed);
       if (!allowPrivate && !isPublicUrl(trimmed)) {
         return apiError(
-          "Base URL points to an internal/private host. Ops must allow it on this instance via ALLOW_LOCAL_AI_PRIVATE_HOSTS — set it to the exact host (e.g. ollama.lan) or to true for any private host (intended for self-hosted Ollama / LM Studio).",
+          "Base URL points to an internal/private host. The operator must allow its exact origin on this instance via AI_PRIVATE_ORIGINS (e.g. http://ollama.lan:11434) for a self-hosted Ollama / LM Studio.",
           422,
         );
       }
@@ -228,7 +228,7 @@ export const PATCH = apiHandler(async (request: NextRequest) => {
       // here as much as it does for Ollama.
       if (!isLocalAiHostAllowed(trimmed) && !isPublicUrl(trimmed)) {
         return apiError(
-          "Base URL points to an internal/private host. Ops must allow it on this instance via ALLOW_LOCAL_AI_PRIVATE_HOSTS — set it to the exact host (e.g. litellm.lan) or to true for any private host (intended for a self-hosted gateway).",
+          "Base URL points to an internal/private host. The operator must allow its exact origin on this instance via AI_PRIVATE_ORIGINS (e.g. http://litellm.lan:4000) for a self-hosted gateway.",
           422,
         );
       }
diff --git a/src/lib/__tests__/ssrf-pin-inventory.test.ts b/src/lib/__tests__/ssrf-pin-inventory.test.ts
index 21e4886db..23a5c7934 100644
--- a/src/lib/__tests__/ssrf-pin-inventory.test.ts
+++ b/src/lib/__tests__/ssrf-pin-inventory.test.ts
@@ -9,14 +9,13 @@
  * instead of shipping silently.
  *
  * The LOCAL AI client and (since v1.37.30) the openai-client's gateway and
- * admin-key tags are the deliberate exceptions: they stay CONDITIONAL
- * (`requirePublicHost: !allowPrivate`) so an operator can opt into LAN hosts
- * via `ALLOW_LOCAL_AI_PRIVATE_HOSTS`. v1.18.7 (SECURITY LOW) — that flag is
- * now a host ALLOWLIST (`true` = any private host; a comma-separated list =
- * only those), resolved by `isLocalAiHostAllowed`. We assert the client keeps
- * the conditional shape AND derives `allowPrivate` from the allowlist helper
- * (not a raw `=== "true"` binary), never an unconditional `true` (which would
- * break LAN local models) and never absent (which would re-open the default).
+ * admin-key tags are the deliberate exceptions: an operator can grant a LAN
+ * endpoint. Since v1.39.3 the grant is an exact origin (`AI_PRIVATE_ORIGINS`,
+ * or the legacy host list) and the call is dialled through the pinned
+ * operator-approved dispatcher, never unpinned; the whole policy comes from
+ * `aiEgressPolicyFor`. We assert every AI dial site spreads that policy, and
+ * that neither an unconditional `false` (no pin) nor a local
+ * `isLocalAiHostAllowed` decision (a grant with no pin) reappears there.
  */
 import { readFileSync } from "node:fs";
 import { join } from "node:path";
@@ -35,8 +34,9 @@ describe("SSRF requirePublicHost pin inventory", () => {
   it("openai-client routes person-typed base URLs through the allowlist with a pinned floor", () => {
     const src = read("ai/openai-client.ts");
     expect(src).toMatch(
-      /requirePublicHost:\s*this\.isGateway \|\| this\.type === "admin-key"\s*\? requirePublicHostFor\(this\.config\.baseUrl\)\s*:\s*true/,
+      /this\.isGateway \|\| this\.type === "admin-key"\s*\?\s*aiEgressPolicyFor\(url\)\s*:\s*\{\s*requirePublicHost:\s*true as const\s*\}/,
     );
+    expect(src).not.toMatch(/requirePublicHost:\s*false/);
   });
 
   it("anthropic-client pins the BYO base-URL outbound unconditionally", () => {
@@ -83,16 +83,18 @@ describe("SSRF requirePublicHost pin inventory", () => {
     },
   );
 
-  it("local AI client keeps the pin CONDITIONAL (LAN escape hatch)", () => {
+  it("local AI client dials every call through the shared egress policy", () => {
     const src = read("ai/local-client.ts");
-    // Must carry the conditional form …
-    expect(src).toMatch(/requirePublicHost:\s*!?\s*allowPrivate/);
-    // … and must NOT be hardened to an unconditional true (that would break
-    // a deliberately-private LAN local model).
-    expect(src).not.toMatch(/requirePublicHost:\s*true/);
-    // … and `allowPrivate` must come from the host-allowlist helper, not a
-    // raw binary `=== "true"` (v1.18.7 — the flag became an allowlist).
-    expect(src).toMatch(/allowPrivate\s*=\s*isLocalAiHostAllowed\(/);
-    expect(src).not.toMatch(/ALLOW_LOCAL_AI_PRIVATE_HOSTS\s*===\s*"true"/);
+    // Both the buffered and the streaming call derive the policy …
+    expect([
+      ...src.matchAll(/const egress = aiEgressPolicyFor\(url\)/g),
+    ]).toHaveLength(2);
+    // … and spread it into safeFetch …
+    expect([...src.matchAll(/\.\.\.egress,/g)]).toHaveLength(2);
+    // … with no local decision beside it: no hard-coded pin value (true
+    // would break a granted LAN model, false would drop the pin) and no
+    // grant check that could skip the operator-approved dispatcher.
+    expect(src).not.toMatch(/requirePublicHost:\s*(?:true|false|!)/);
+    expect(src).not.toMatch(/isLocalAiHostAllowed\(/);
   });
 });
diff --git a/src/lib/ai/__tests__/local-client-egress.test.ts b/src/lib/ai/__tests__/local-client-egress.test.ts
new file mode 100644
index 000000000..806c24874
--- /dev/null
+++ b/src/lib/ai/__tests__/local-client-egress.test.ts
@@ -0,0 +1,86 @@
+/**
+ * v1.39.3 — the destination policy an AI call is dialled with.
+ *
+ * A granted private origin goes to `safeFetch` as an exact
+ * `operatorApprovedPrivateOrigin`, which routes the connect through the pinned
+ * operator-approved dispatcher: redirects forbidden, metadata, link-local and
+ * the unspecified address dropped at dial time even for a granted name. Before
+ * this release a granted host was dialled with no pin at all. Everything else
+ * is pinned to public hosts.
+ */
+import { afterEach, beforeEach, describe, expect, it, vi } from "vitest";
+
+const { safeFetch } = vi.hoisted(() => ({ safeFetch: vi.fn() }));
+vi.mock("@/lib/safe-fetch", async (importOriginal) => ({
+  ...(await importOriginal<typeof import("@/lib/safe-fetch")>()),
+  safeFetch,
+}));
+
+import { LocalOpenAICompatibleClient } from "../local-client";
+import { _resetAiGrantsForTests } from "../local-host-allowlist";
+import { singleUserTurn } from "../types";
+
+function client(baseUrl: string) {
+  return new LocalOpenAICompatibleClient({
+    apiKey: null,
+    model: "llama3:8b",
+    baseUrl,
+  });
+}
+
+beforeEach(() => {
+  _resetAiGrantsForTests();
+  safeFetch.mockReset();
+  safeFetch.mockResolvedValue(
+    new Response(
+      JSON.stringify({ choices: [{ message: { content: "ok" } }] }),
+      { status: 200, headers: { "content-type": "application/json" } },
+    ),
+  );
+  vi.stubEnv("AI_PRIVATE_ORIGINS", "");
+  vi.stubEnv("ALLOW_LOCAL_AI_PRIVATE_HOSTS", "");
+  vi.spyOn(console, "warn").mockImplementation(() => {});
+});
+
+afterEach(() => {
+  vi.unstubAllEnvs();
+  vi.restoreAllMocks();
+});
+
+const call = singleUserTurn({ system: "sys", user: "hi" });
+
+describe("Local AI egress policy", () => {
+  it("dials a granted origin through the operator-approved pin", async () => {
+    vi.stubEnv("AI_PRIVATE_ORIGINS", "http://ollama.lan:11434");
+    await client("http://ollama.lan:11434/v1").generateCompletion(call);
+    const opts = safeFetch.mock.calls[0][2];
+    expect(opts).toMatchObject({
+      requirePublicHost: false,
+      operatorApprovedPrivateOrigin: "http://ollama.lan:11434",
+    });
+  });
+
+  it("pins an ungranted private URL to public hosts, so safeFetch refuses it", async () => {
+    await client("http://10.0.0.5:11434/v1").generateCompletion(call);
+    const opts = safeFetch.mock.calls[0][2];
+    expect(opts.requirePublicHost).toBe(true);
+    expect(opts).not.toHaveProperty("operatorApprovedPrivateOrigin");
+  });
+
+  it("the retired `=true` grants nothing any more", async () => {
+    vi.stubEnv("ALLOW_LOCAL_AI_PRIVATE_HOSTS", "true");
+    await client("http://10.0.0.5:11434/v1").generateCompletion(call);
+    expect(safeFetch.mock.calls[0][2].requirePublicHost).toBe(true);
+  });
+
+  it("streams through the same policy", async () => {
+    vi.stubEnv("ALLOW_LOCAL_AI_PRIVATE_HOSTS", "ollama.lan");
+    await client("http://ollama.lan:11434/v1")
+      .generateCompletionStream(call, () => {})
+      .catch(() => undefined);
+    expect(safeFetch.mock.calls[0][2]).toMatchObject({
+      requirePublicHost: false,
+      operatorApprovedPrivateOrigin: "http://ollama.lan:11434",
+    });
+  });
+});
diff --git a/src/lib/ai/__tests__/local-client.test.ts b/src/lib/ai/__tests__/local-client.test.ts
index b75148df3..739540f0e 100644
--- a/src/lib/ai/__tests__/local-client.test.ts
+++ b/src/lib/ai/__tests__/local-client.test.ts
@@ -1,4 +1,11 @@
 import { describe, it, expect, vi, beforeEach, afterEach } from "vitest";
+
+// The client's destination policy has its own suite
+// (`local-host-allowlist.test.ts`, `local-client-egress.test.ts`). Here the
+// grant is stubbed to "no policy" so the request shape can be asserted
+// against the stubbed global fetch; the pinned dispatchers drive undici's own
+// fetch, which a global stub cannot see.
+vi.mock("../local-host-allowlist", () => ({ aiEgressPolicyFor: () => ({}) }));
 import {
   LocalOpenAICompatibleClient,
   resetLocalJsonDialectCache,
@@ -8,10 +15,6 @@ import { singleUserTurn } from "../types";
 describe("LocalOpenAICompatibleClient", () => {
   beforeEach(() => {
     vi.restoreAllMocks();
-    // These tests cover the operator-opt-in path (LM Studio / Ollama on
-    // an RFC1918 host). The safeFetch wrapper enforces the SSRF guard
-    // unless the operator explicitly accepts a private host.
-    vi.stubEnv("ALLOW_LOCAL_AI_PRIVATE_HOSTS", "true");
     // v1.28.28 (#470) — the JSON dialect is cached per baseUrl for the
     // process; clear it so each test starts from the default dialect.
     resetLocalJsonDialectCache();
diff --git a/src/lib/ai/__tests__/local-host-allowlist.test.ts b/src/lib/ai/__tests__/local-host-allowlist.test.ts
index 0341767aa..62cf568ef 100644
--- a/src/lib/ai/__tests__/local-host-allowlist.test.ts
+++ b/src/lib/ai/__tests__/local-host-allowlist.test.ts
@@ -1,57 +1,116 @@
-import { afterEach, describe, expect, it, vi } from "vitest";
+import { afterEach, beforeEach, describe, expect, it, vi } from "vitest";
 
 import {
+  _resetAiGrantsForTests,
+  aiEgressPolicyFor,
+  grantedAiOrigin,
   isLocalAiHostAllowed,
-  requirePublicHostFor,
+  legacyAnyHostConfigured,
+  LEGACY_ANY_HOST_WARNING,
 } from "../local-host-allowlist";
 
+/**
+ * v1.39.3 — AI base URLs on a private network are granted by exact origin
+ * (`AI_PRIVATE_ORIGINS`), the grammar the notification and Nightscout grants
+ * use, and dialled through the pinned operator-approved dispatcher. The legacy
+ * host list keeps working; its `=true` form grants nothing.
+ */
+
+let warn: ReturnType<typeof vi.spyOn>;
+
+beforeEach(() => {
+  _resetAiGrantsForTests();
+  vi.stubEnv("AI_PRIVATE_ORIGINS", "");
+  vi.stubEnv("ALLOW_LOCAL_AI_PRIVATE_HOSTS", "");
+  warn = vi.spyOn(console, "warn").mockImplementation(() => {});
+});
+
 afterEach(() => {
   vi.unstubAllEnvs();
+  warn.mockRestore();
 });
 
-describe("isLocalAiHostAllowed", () => {
-  it("denies every private host when unset (secure default)", () => {
-    vi.stubEnv("ALLOW_LOCAL_AI_PRIVATE_HOSTS", "");
+describe("AI_PRIVATE_ORIGINS", () => {
+  it("grants nothing when unset", () => {
     expect(isLocalAiHostAllowed("http://10.0.0.5:11434/v1")).toBe(false);
     expect(isLocalAiHostAllowed("http://ollama.lan/v1")).toBe(false);
   });
 
-  it('treats "false" as denying every private host', () => {
-    vi.stubEnv("ALLOW_LOCAL_AI_PRIVATE_HOSTS", "false");
-    expect(isLocalAiHostAllowed("http://10.0.0.5/v1")).toBe(false);
+  it("grants exactly the listed origin, not its siblings", () => {
+    vi.stubEnv(
+      "AI_PRIVATE_ORIGINS",
+      "http://ollama.lan:11434, http://10.0.0.5:4000",
+    );
+    expect(grantedAiOrigin("http://ollama.lan:11434/v1/chat")).toBe(
+      "http://ollama.lan:11434",
+    );
+    expect(isLocalAiHostAllowed("http://10.0.0.5:4000/v1")).toBe(true);
+    // Another port, another scheme, a subdomain: all separate origins.
+    expect(isLocalAiHostAllowed("http://ollama.lan:8080/v1")).toBe(false);
+    expect(isLocalAiHostAllowed("https://ollama.lan:11434/v1")).toBe(false);
+    expect(isLocalAiHostAllowed("http://evil.ollama.lan:11434/v1")).toBe(false);
+    expect(isLocalAiHostAllowed("http://10.0.0.6:4000/v1")).toBe(false);
   });
 
-  it('keeps the legacy "true" = any private host', () => {
-    vi.stubEnv("ALLOW_LOCAL_AI_PRIVATE_HOSTS", "true");
-    expect(isLocalAiHostAllowed("http://10.0.0.5:11434/v1")).toBe(true);
-    expect(isLocalAiHostAllowed("http://anything.lan/v1")).toBe(true);
-  });
-
-  it("permits ONLY the listed hosts when given a comma-separated allowlist", () => {
-    vi.stubEnv("ALLOW_LOCAL_AI_PRIVATE_HOSTS", "ollama.lan, 10.0.0.5");
-    expect(isLocalAiHostAllowed("http://ollama.lan:11434/v1")).toBe(true);
-    expect(isLocalAiHostAllowed("http://10.0.0.5:11434/v1")).toBe(true);
-    // A different private host is still rejected — the metadata endpoint and
-    // internal panels no longer ride the binary flag.
+  it("never grants metadata, link-local or the unspecified address, even when listed", () => {
+    vi.stubEnv(
+      "AI_PRIVATE_ORIGINS",
+      "http://169.254.169.254,http://[fe80::1]:80,http://0.0.0.0:11434",
+    );
     expect(isLocalAiHostAllowed("http://169.254.169.254/latest/")).toBe(false);
+    expect(isLocalAiHostAllowed("http://[fe80::1]/v1")).toBe(false);
+    expect(isLocalAiHostAllowed("http://0.0.0.0:11434/v1")).toBe(false);
+    expect(warn).toHaveBeenCalled();
+  });
+});
+
+describe("ALLOW_LOCAL_AI_PRIVATE_HOSTS (legacy)", () => {
+  it("still grants a listed host on any port", () => {
+    vi.stubEnv("ALLOW_LOCAL_AI_PRIVATE_HOSTS", " Ollama.LAN , 10.0.0.5 ");
+    expect(isLocalAiHostAllowed("http://ollama.lan:11434/v1")).toBe(true);
+    expect(isLocalAiHostAllowed("http://10.0.0.5:8080/v1")).toBe(true);
     expect(isLocalAiHostAllowed("http://10.0.0.6/v1")).toBe(false);
   });
 
-  it("matches hostnames case-insensitively and ignores blank entries", () => {
-    vi.stubEnv("ALLOW_LOCAL_AI_PRIVATE_HOSTS", " , Ollama.LAN , ");
-    expect(isLocalAiHostAllowed("http://ollama.lan/v1")).toBe(true);
+  it("no longer opens every private host with `true`, and says so", () => {
+    vi.stubEnv("ALLOW_LOCAL_AI_PRIVATE_HOSTS", "true");
+    expect(isLocalAiHostAllowed("http://10.0.0.5:11434/v1")).toBe(false);
+    expect(isLocalAiHostAllowed("http://169.254.169.254/latest/")).toBe(false);
+    expect(warn).toHaveBeenCalledWith(LEGACY_ANY_HOST_WARNING);
+    expect(legacyAnyHostConfigured()).toBe(true);
   });
 
-  it("never permits an unparseable URL", () => {
-    vi.stubEnv("ALLOW_LOCAL_AI_PRIVATE_HOSTS", "true");
-    expect(isLocalAiHostAllowed("not a url")).toBe(true); // host-agnostic "any"
+  it("refuses a metadata literal and a host:port entry instead of granting a guess", () => {
+    vi.stubEnv(
+      "ALLOW_LOCAL_AI_PRIVATE_HOSTS",
+      "169.254.169.254,ollama.lan:11434",
+    );
+    expect(isLocalAiHostAllowed("http://169.254.169.254/latest/")).toBe(false);
+    expect(isLocalAiHostAllowed("http://ollama.lan:11434/v1")).toBe(false);
+    expect(warn).toHaveBeenCalledTimes(2);
+  });
+
+  it("an unparseable URL is never granted", () => {
     vi.stubEnv("ALLOW_LOCAL_AI_PRIVATE_HOSTS", "ollama.lan");
     expect(isLocalAiHostAllowed("not a url")).toBe(false);
   });
+});
 
-  it("requirePublicHostFor is the inverse of the allowlist", () => {
-    vi.stubEnv("ALLOW_LOCAL_AI_PRIVATE_HOSTS", "ollama.lan");
-    expect(requirePublicHostFor("http://ollama.lan/v1")).toBe(false);
-    expect(requirePublicHostFor("http://10.0.0.6/v1")).toBe(true);
+describe("aiEgressPolicyFor", () => {
+  it("pins an ungranted URL to the public dispatcher", () => {
+    expect(aiEgressPolicyFor("http://10.0.0.5/v1")).toEqual({
+      requirePublicHost: true,
+    });
+    expect(aiEgressPolicyFor("https://api.example.com/v1")).toEqual({
+      requirePublicHost: true,
+    });
+  });
+
+  it("hands a granted URL to the operator-approved dispatcher for exactly its origin", () => {
+    vi.stubEnv("AI_PRIVATE_ORIGINS", "http://ollama.lan:11434");
+    expect(aiEgressPolicyFor("http://ollama.lan:11434/v1/chat")).toEqual({
+      requirePublicHost: false,
+      operatorApprovedPrivateOrigin: "http://ollama.lan:11434",
+    });
   });
 });
diff --git a/src/lib/ai/__tests__/provider-contract.test.ts b/src/lib/ai/__tests__/provider-contract.test.ts
index 39ed56e34..70d993be6 100644
--- a/src/lib/ai/__tests__/provider-contract.test.ts
+++ b/src/lib/ai/__tests__/provider-contract.test.ts
@@ -53,8 +53,8 @@ describe("AIProvider contract", () => {
     vi.restoreAllMocks();
     // The local-provider contract test points at a localhost OpenAI-
     // compatible server (Ollama / LM Studio). safeFetch enforces the
-    // SSRF guard unless the operator explicitly opts in.
-    vi.stubEnv("ALLOW_LOCAL_AI_PRIVATE_HOSTS", "true");
+    // SSRF guard unless the operator grants that exact origin.
+    vi.stubEnv("AI_PRIVATE_ORIGINS", "http://localhost:11434");
   });
 
   afterEach(() => {
diff --git a/src/lib/ai/local-client.ts b/src/lib/ai/local-client.ts
index cd62b7dd9..6d38b8d5d 100644
--- a/src/lib/ai/local-client.ts
+++ b/src/lib/ai/local-client.ts
@@ -1,6 +1,6 @@
 import { safeFetch } from "@/lib/safe-fetch";
 import { annotate } from "@/lib/logging/context";
-import { isLocalAiHostAllowed } from "./local-host-allowlist";
+import { aiEgressPolicyFor } from "./local-host-allowlist";
 import type {
   AIProvider,
   AiMessage,
@@ -193,11 +193,12 @@ export class LocalOpenAICompatibleClient implements AIProvider {
     // pin (issue #217) extends the same flag with a connect-time
     // resolved-IP check to also defeat DNS rebinding. Operators who
     // legitimately point at a self-hosted Ollama / LM Studio on an
-    // RFC1918 address opt in via `ALLOW_LOCAL_AI_PRIVATE_HOSTS` — either the
-    // legacy `=true` (any private host) or a comma-separated host allowlist
-    // (only those hostnames). Enforced at write-time in /api/user/ai-provider
-    // too. v1.18.7 (SECURITY LOW) — the binary flag became a host allowlist.
-    const allowPrivate = isLocalAiHostAllowed(url);
+    // RFC1918 address opt in via `AI_PRIVATE_ORIGINS` (exact origins) or the
+    // legacy host list. A granted origin is dialled through the pinned
+    // operator-approved dispatcher, never unpinned, so the metadata range and
+    // link-local stay unreachable even for a granted name (v1.39.3). Enforced
+    // at write-time in /api/user/ai-provider too.
+    const egress = aiEgressPolicyFor(url);
     const res = await safeFetch(
       url,
       {
@@ -210,7 +211,7 @@ export class LocalOpenAICompatibleClient implements AIProvider {
       {
         // v1.22 (#89) — honour the caller's per-request timeout override; default 60 s.
         timeoutMs: params.timeoutMs ?? 60_000,
-        requirePublicHost: !allowPrivate,
+        ...egress,
         signal: params.signal,
       },
     );
@@ -335,7 +336,7 @@ export class LocalOpenAICompatibleClient implements AIProvider {
     onDelta: (delta: string) => void,
   ): Promise<CompletionResult> {
     const { url, headers, body } = this.buildRequest(params, true);
-    const allowPrivate = isLocalAiHostAllowed(url);
+    const egress = aiEgressPolicyFor(url);
 
     // Idle controller: aborts the upstream when no chunk arrives within the
     // per-idle window. Composed with the caller's cancel signal so a client
@@ -355,7 +356,7 @@ export class LocalOpenAICompatibleClient implements AIProvider {
           // Absolute backstop only — the per-idle timer below is the real
           // ceiling. A server that streams forever still cannot pin a worker.
           timeoutMs: STREAM_ABSOLUTE_CEILING_MS,
-          requirePublicHost: !allowPrivate,
+          ...egress,
           signal,
         },
       );
diff --git a/src/lib/ai/local-host-allowlist.ts b/src/lib/ai/local-host-allowlist.ts
index bf26845f0..bfa13c69c 100644
--- a/src/lib/ai/local-host-allowlist.ts
+++ b/src/lib/ai/local-host-allowlist.ts
@@ -1,76 +1,165 @@
 /**
- * v1.18.7 (SECURITY LOW) — host allowlist for local/private AI endpoints.
+ * Which private hosts an AI base URL may reach, and how the call is dialled.
  *
- * The operator escape-hatch for pointing the server at a self-hosted
- * Ollama / LM Studio on an RFC1918 address used to be a single binary flag,
- * `ALLOW_LOCAL_AI_PRIVATE_HOSTS=true`, which opened EVERY private/internal
- * host at once — including the cloud-metadata endpoint and any internal admin
- * panel. This narrows it to an explicit allowlist while keeping the binary
- * `true` for backward compatibility:
+ * A base URL for the Local provider, the OpenAI-compatible gateway and the
+ * operator's admin key is typed by a person, and for the first two that
+ * person is any user of the instance. By default only public hosts are
+ * reachable. An operator who runs Ollama, LM Studio or a LiteLLM gateway on
+ * their own network grants it through the environment, the same exact-origin
+ * grammar `NOTIFICATION_PRIVATE_ORIGINS` and `NIGHTSCOUT_PRIVATE_ORIGINS` use
+ * (`src/lib/private-origin-policy.ts`):
  *
- *   - unset / "" / "false"      → no private host allowed (the secure default).
- *   - "true"                    → ANY private host allowed (legacy behaviour).
- *   - "ollama.lan, 10.0.0.5"    → ONLY those exact hostnames allowed; every
- *                                 other private host is still rejected.
+ *   AI_PRIVATE_ORIGINS="http://ollama.lan:11434,http://10.0.0.5:4000"
  *
- * This is layered ON TOP of the `safeFetch` / `isPublicUrl` floor: a public
- * URL never needs the allowlist, and a private host not on the list is
- * rejected exactly as before. The allowlist only ever WIDENS what the
- * operator deliberately permits, never narrows the public-host floor.
+ * A grant is one `scheme://host[:port]`. It is never a wildcard, suffix or
+ * range. A granted call is dialled through the pinned operator-approved
+ * dispatcher with redirects forbidden, so a listed name that resolves to the
+ * metadata range, link-local or the unspecified address is dropped at connect
+ * time, and a literal in those ranges cannot be granted at all. Nothing in a
+ * request or a settings field can confer a grant.
+ *
+ * `ALLOW_LOCAL_AI_PRIVATE_HOSTS` (v1.18.7) stays readable for the one form that
+ * can be made safe: a comma-separated host list (`ollama.lan,10.0.0.5`) still
+ * grants those exact hosts, on any port, dialled the same pinned way. Its
+ * `=true` form granted every private host on the network to every user and
+ * cannot be made safe, so from v1.39.3 it grants nothing; the boot readiness
+ * summary and a one-time warning say so and name the replacement.
  */
+import { isIP } from "node:net";
 
-const ENV_VAR = "ALLOW_LOCAL_AI_PRIVATE_HOSTS";
+import {
+  isNeverGrantableHost,
+  originOfUrl,
+  parsePrivateOrigins,
+} from "@/lib/private-origin-policy";
 
-/** Parsed allowlist policy from the env var. */
-type LocalHostPolicy =
-  | { kind: "none" }
-  | { kind: "any" }
-  | { kind: "hosts"; hosts: ReadonlySet<string> };
+const ORIGINS_ENV = "AI_PRIVATE_ORIGINS";
+const LEGACY_ENV = "ALLOW_LOCAL_AI_PRIVATE_HOSTS";
 
-function parsePolicy(): LocalHostPolicy {
-  const raw = process.env[ENV_VAR]?.trim();
-  if (!raw || raw.toLowerCase() === "false") return { kind: "none" };
-  if (raw.toLowerCase() === "true") return { kind: "any" };
-  const hosts = raw
-    .split(",")
-    .map((h) => h.trim().toLowerCase())
-    .filter((h) => h.length > 0);
-  if (hosts.length === 0) return { kind: "none" };
-  return { kind: "hosts", hosts: new Set(hosts) };
+/** The operator-facing sentence for a legacy `=true`. Shared with the boot summary. */
+export const LEGACY_ANY_HOST_WARNING =
+  `${LEGACY_ENV}=true no longer opens every private host (it let any user ` +
+  `reach any service on the operator's network). List the AI endpoint ` +
+  `instead: ${ORIGINS_ENV}="http://ollama.lan:11434".`;
+
+interface Grants {
+  origins: ReadonlySet<string>;
+  legacyHosts: ReadonlySet<string>;
 }
 
-/** Extract the lowercased hostname from a URL string, or null if unparseable. */
-function hostnameOf(url: string): string | null {
-  try {
-    return new URL(url).hostname.toLowerCase();
-  } catch {
-    return null;
+let cachedKey: string | null = null;
+let cached: Grants = { origins: new Set(), legacyHosts: new Set() };
+const warned = new Set<string>();
+
+function warnOnce(message: string): void {
+  if (warned.has(message)) return;
+  warned.add(message);
+  console.warn(message);
+}
+
+/** True when the legacy variable carries its retired `=true` form. */
+export function legacyAnyHostConfigured(
+  env: Record<string, string | undefined> = process.env,
+): boolean {
+  return env[LEGACY_ENV]?.trim().toLowerCase() === "true";
+}
+
+function parseLegacyHosts(raw: string | undefined): ReadonlySet<string> {
+  const value = raw?.trim() ?? "";
+  if (!value || value.toLowerCase() === "false") return new Set();
+  if (value.toLowerCase() === "true") {
+    warnOnce(LEGACY_ANY_HOST_WARNING);
+    return new Set();
   }
+  const hosts = new Set<string>();
+  for (const entry of value.split(",")) {
+    const host = entry.trim().toLowerCase();
+    if (!host) continue;
+    // A host, not a URL: anything with a scheme, path, port or wildcard is a
+    // typo for the origin list, and granting a guess would widen it.
+    const bare = host.replace(/^\[|\]$/g, "");
+    const isIpv6 = isIP(bare) === 6;
+    if (!/^[a-z0-9.\-[\]:]+$/.test(host) || (host.includes(":") && !isIpv6)) {
+      warnOnce(
+        `${LEGACY_ENV}: ignoring entry "${host.slice(0, 200)}": a host name or address, nothing else; use ${ORIGINS_ENV} for a full origin`,
+      );
+      continue;
+    }
+    if (isNeverGrantableHost(host)) {
+      warnOnce(
+        `${LEGACY_ENV}: ignoring entry "${host}": the unspecified address, link-local and metadata addresses cannot be granted`,
+      );
+      continue;
+    }
+    hosts.add(bare);
+  }
+  return hosts;
+}
+
+function grants(): Grants {
+  const originsRaw = process.env.AI_PRIVATE_ORIGINS;
+  const legacyRaw = process.env.ALLOW_LOCAL_AI_PRIVATE_HOSTS;
+  const key = `${originsRaw ?? ""}\u0000${legacyRaw ?? ""}`;
+  if (key === cachedKey) return cached;
+  cached = {
+    origins: parsePrivateOrigins(originsRaw, (entry, reason) => {
+      warnOnce(`${ORIGINS_ENV}: ignoring entry "${entry}": ${reason}`);
+    }),
+    legacyHosts: parseLegacyHosts(legacyRaw),
+  };
+  cachedKey = key;
+  return cached;
 }
 
 /**
- * True when the operator has explicitly permitted this (private) host for a
- * local AI endpoint. `"true"` permits any host (legacy); a host list permits
- * only listed hostnames. An unparseable URL is never permitted.
- *
- * NOTE: this answers "did the operator opt this host in?", NOT "is this host
- * public?". Callers still run the `isPublicUrl` / `safeFetch` floor; this is
- * the private-host override on top of it.
+ * The granted origin for this URL, or null. A URL is granted when its exact
+ * origin is listed in `AI_PRIVATE_ORIGINS`, or its host is listed in the
+ * legacy host list. A host no grant may name is never granted, whatever the
+ * lists say.
+ */
+export function grantedAiOrigin(url: string): string | null {
+  // Parse the grants first, so a malformed entry is reported on the first
+  // call that could have used it rather than never.
+  const { origins, legacyHosts } = grants();
+  const origin = originOfUrl(url);
+  if (!origin) return null;
+  const hostname = new URL(origin).hostname.toLowerCase();
+  // Both lists refuse a never-grantable literal when they are parsed, so
+  // neither can hold one; a listed name that resolves into those ranges is
+  // dropped by the pinned dispatcher at dial time.
+  if (origins.has(origin)) return origin;
+  if (legacyHosts.has(hostname.replace(/^\[|\]$/g, ""))) return origin;
+  return null;
+}
+
+/**
+ * True when the operator granted this URL's origin. Answers "may this private
+ * host be saved and called?", never "is it public?": callers still accept a
+ * public URL without any grant.
  */
 export function isLocalAiHostAllowed(url: string): boolean {
-  const policy = parsePolicy();
-  if (policy.kind === "none") return false;
-  if (policy.kind === "any") return true;
-  const host = hostnameOf(url);
-  return host !== null && policy.hosts.has(host);
+  return grantedAiOrigin(url) !== null;
 }
 
 /**
- * Whether `safeFetch` must require a public host for this URL: true unless the
- * operator has opted this private host into the allowlist. Kept as a thin
- * wrapper so the call sites read `requirePublicHost: requirePublicHostFor(url)`
- * and the conditional intent stays obvious.
+ * The `safeFetch` destination policy for an AI call to `url`: the pinned
+ * public dispatcher unless the operator granted the origin, and then the
+ * pinned operator-approved dispatcher for that one origin. Spread into the
+ * `safeFetch` options.
  */
-export function requirePublicHostFor(url: string): boolean {
-  return !isLocalAiHostAllowed(url);
+export function aiEgressPolicyFor(
+  url: string,
+):
+  | { requirePublicHost: true }
+  | { requirePublicHost: false; operatorApprovedPrivateOrigin: string } {
+  const origin = grantedAiOrigin(url);
+  return origin === null
+    ? { requirePublicHost: true }
+    : { requirePublicHost: false, operatorApprovedPrivateOrigin: origin };
+}
+
+/** Test helper: forget the parsed grants and the warnings already printed. */
+export function _resetAiGrantsForTests(): void {
+  cachedKey = null;
+  warned.clear();
 }
diff --git a/src/lib/ai/openai-client.ts b/src/lib/ai/openai-client.ts
index 240e70ec9..409ba5d8c 100644
--- a/src/lib/ai/openai-client.ts
+++ b/src/lib/ai/openai-client.ts
@@ -8,7 +8,7 @@ import {
   jsonModeDialectFor,
   rememberJsonModeDialect,
 } from "./json-dialect";
-import { requirePublicHostFor } from "./local-host-allowlist";
+import { aiEgressPolicyFor } from "./local-host-allowlist";
 import {
   buildOpenAIMessages,
   buildOpenAITools,
@@ -152,7 +152,8 @@ export class OpenAIClient implements AIProvider {
       // v1.21.5 — honour the caller's per-request timeout override; default 60 s.
       // v1.33.1 (#470) — the gateway tag reuses the Local provider's host
       // policy verbatim: public hosts always; a private host only when the
-      // operator allowlisted it via ALLOW_LOCAL_AI_PRIVATE_HOSTS.
+      // operator granted its origin (`AI_PRIVATE_ORIGINS` or the legacy host
+      // list), and then only through the pinned operator-approved dispatcher.
       //
       // v1.37.30 — the `admin-key` tag consults the same policy. Its base URL
       // is the one other operator-typed URL in this client (an operator can
@@ -167,10 +168,9 @@ export class OpenAIClient implements AIProvider {
       // operator legitimately redirects.
       {
         timeoutMs: params.timeoutMs ?? 60_000,
-        requirePublicHost:
-          this.isGateway || this.type === "admin-key"
-            ? requirePublicHostFor(this.config.baseUrl)
-            : true,
+        ...(this.isGateway || this.type === "admin-key"
+          ? aiEgressPolicyFor(url)
+          : { requirePublicHost: true as const }),
         signal: params.signal,
       },
     );
diff --git a/src/lib/ai/provider.ts b/src/lib/ai/provider.ts
index 7586504b2..ab16974bf 100644
--- a/src/lib/ai/provider.ts
+++ b/src/lib/ai/provider.ts
@@ -1189,8 +1189,8 @@ export async function resolveProviderForTest(
       if (!baseUrl) {
         throw new AITestConfigError(422, "Local provider requires a base URL");
       }
-      // v1.18.7 (SECURITY LOW) — host allowlist (`true` = any private host;
-      // a comma-separated host list = only those) replaces the binary flag.
+      // A private host only when the operator granted its origin
+      // (`AI_PRIVATE_ORIGINS` or the legacy host list; v1.39.3).
       const allowPrivate = isLocalAiHostAllowed(baseUrl);
       if (!allowPrivate && !isPublicUrl(baseUrl)) {
         throw new AITestConfigError(
diff --git a/src/lib/boot/__tests__/readiness-summary.test.ts b/src/lib/boot/__tests__/readiness-summary.test.ts
index 4d6211ba2..fc23b92ee 100644
--- a/src/lib/boot/__tests__/readiness-summary.test.ts
+++ b/src/lib/boot/__tests__/readiness-summary.test.ts
@@ -129,3 +129,23 @@ describe("formatReadiness + logReadinessSummary", () => {
     expect(errors).toHaveLength(1);
   });
 });
+
+describe("collectReadiness — retired AI grant", () => {
+  it("warns, without blocking, when ALLOW_LOCAL_AI_PRIVATE_HOSTS is still `true`", () => {
+    const env = { ...baseEnv(), ALLOW_LOCAL_AI_PRIVATE_HOSTS: " TRUE " };
+    const line = lineFor(env, "ALLOW_LOCAL_AI_PRIVATE_HOSTS");
+    expect(line?.status).toBe("warn");
+    expect(line?.detail).toContain("AI_PRIVATE_ORIGINS");
+    expect(collectReadiness(env).hasBlocker).toBe(false);
+  });
+
+  it("says nothing for a host list or an unset variable", () => {
+    expect(
+      lineFor(
+        { ...baseEnv(), ALLOW_LOCAL_AI_PRIVATE_HOSTS: "ollama.lan" },
+        "ALLOW_LOCAL_AI_PRIVATE_HOSTS",
+      ),
+    ).toBeUndefined();
+    expect(lineFor(baseEnv(), "ALLOW_LOCAL_AI_PRIVATE_HOSTS")).toBeUndefined();
+  });
+});
diff --git a/src/lib/boot/readiness-summary.ts b/src/lib/boot/readiness-summary.ts
index 6f1e79f15..58869d528 100644
--- a/src/lib/boot/readiness-summary.ts
+++ b/src/lib/boot/readiness-summary.ts
@@ -15,6 +15,11 @@
  * signpost in front of them.
  */
 
+import {
+  LEGACY_ANY_HOST_WARNING,
+  legacyAnyHostConfigured,
+} from "@/lib/ai/local-host-allowlist";
+
 export interface ReadinessLine {
   /** Short label, e.g. "ENCRYPTION_KEY". */
   label: string;
@@ -121,6 +126,19 @@ export function collectReadiness(env: Env = process.env): ReadinessReport {
       : "Secure cookie flag OFF — fine for plain-HTTP LAN; do NOT expose plain HTTP to the internet",
   });
 
+  // --- Retired AI grant (a warning, never a blocker) ----------------------
+  // `ALLOW_LOCAL_AI_PRIVATE_HOSTS=true` used to open every private host on
+  // the network to every user's AI base URL. It grants nothing since
+  // v1.39.3; an operator who still has it set has a local AI endpoint that
+  // just stopped answering, and this is the line that says why.
+  if (legacyAnyHostConfigured(env)) {
+    lines.push({
+      label: "ALLOW_LOCAL_AI_PRIVATE_HOSTS",
+      status: "warn",
+      detail: LEGACY_ANY_HOST_WARNING,
+    });
+  }
+
   // --- Optional subsystems (informational) --------------------------------
   const optional: Array<[string, boolean]> = [
     ["SMTP (email)", allSet(env, ["SMTP_HOST", "SMTP_PORT", "SMTP_FROM"])],