diff --git a/backend/apps/agents/tools/url_guard.py b/backend/apps/agents/tools/url_guard.py
new file mode 100644
index 000000000..6253f223f
--- /dev/null
+++ b/backend/apps/agents/tools/url_guard.py
@@ -0,0 +1,93 @@
+"""SSRF guard for the WebFetch tool.
+
+The agent picks which URL to fetch, and that choice can be steered by
+content on pages it already read (prompt injection). So every fetch URL,
+and every redirect hop, is treated as hostile: only http(s), and never a
+host that resolves to a private, loopback, link-local, or cloud-metadata
+address. Redirects are followed manually so a public URL can't 30x-pivot
+into the LAN behind our back.
+"""
+
+from __future__ import annotations
+
+import ipaddress
+import socket
+from urllib.parse import urlsplit
+
+_ALLOWED_SCHEMES = ("http", "https")
+_MAX_REDIRECTS = 5
+
+
+class BlockedURLError(Exception):
+    """Raised when a URL (or a redirect target) points somewhere unsafe."""
+
+
+def _ip_is_blocked(ip: str) -> bool:
+    try:
+        addr = ipaddress.ip_address(ip)
+    except ValueError:
+        return True
+    return (
+        addr.is_private
+        or addr.is_loopback
+        or addr.is_link_local
+        or addr.is_reserved
+        or addr.is_multicast
+        or addr.is_unspecified
+        or not addr.is_global
+    )
+
+
+def _resolve_host(host: str) -> list[str]:
+    """All A/AAAA records for host. A hostname can resolve to several IPs
+    (round-robin, dual-stack); one bad one is enough to block."""
+    try:
+        infos = socket.getaddrinfo(host, None)
+    except socket.gaierror as exc:
+        raise BlockedURLError(f"could not resolve host {host!r}") from exc
+    return [info[4][0] for info in infos]
+
+
+def validate_fetch_url(url: str) -> None:
+    """Reject non-http(s) schemes and hosts that resolve to non-public IPs.
+
+    A bare IP literal is checked directly; a hostname is resolved and every
+    returned address must be public. Raises BlockedURLError otherwise.
+    """
+    parts = urlsplit(url)
+    if parts.scheme.lower() not in _ALLOWED_SCHEMES:
+        raise BlockedURLError(
+            f"url scheme must be http or https (got {parts.scheme or 'none'!r})"
+        )
+    host = parts.hostname
+    if not host:
+        raise BlockedURLError("url has no host")
+
+    # IP literal: check as-is so DNS rebinding can't sneak one past us.
+    try:
+        if _ip_is_blocked(str(ipaddress.ip_address(host))):
+            raise BlockedURLError(f"host {host!r} resolves to a non-public address")
+        return
+    except ValueError:
+        pass
+
+    for ip in _resolve_host(host):
+        if _ip_is_blocked(ip):
+            raise BlockedURLError(f"host {host!r} resolves to a non-public address")
+
+
+async def fetch_guarded(client, url: str):
+    """GET `url` with manual redirect following, validating every hop.
+
+    `client` must be an httpx.AsyncClient created with follow_redirects=False
+    so this layer (not httpx) decides whether each Location is safe.
+    """
+    current = url
+    for _ in range(_MAX_REDIRECTS + 1):
+        validate_fetch_url(current)
+        resp = await client.get(current)
+        if resp.is_redirect and resp.headers.get("location"):
+            current = str(resp.next_request.url) if resp.next_request else resp.headers["location"]
+            continue
+        return resp
+    raise BlockedURLError(f"too many redirects fetching {url}")
diff --git a/backend/apps/agents/tools/web.py b/backend/apps/agents/tools/web.py
index ff039abe5..2959954fb 100644
--- a/backend/apps/agents/tools/web.py
+++ b/backend/apps/agents/tools/web.py
@@ -9,6 +9,7 @@
 import httpx
 
 from backend.apps.agents.tools.base import BaseTool, ToolContext
+from backend.apps.agents.tools.url_guard import BlockedURLError, fetch_guarded
 
 _HTTP_TIMEOUT = 30
 _MAX_OUTPUT_BYTES = 250 * 1024  # ~250 KB covers ~95% of articles/wikis/docs.
@@ -168,13 +169,17 @@ async def execute(self, input_data: dict, context: ToolContext) -> list[dict]:
         prompt: str | None = input_data.get("prompt")
 
         try:
+            # follow_redirects=False so url_guard re-checks every hop; httpx's own
+            # auto-redirect would happily 30x-pivot us into the LAN.
             async with httpx.AsyncClient(
                 timeout=_HTTP_TIMEOUT,
-                follow_redirects=True,
+                follow_redirects=False,
                 headers={"User-Agent": _USER_AGENT},
             ) as client:
-                resp = await client.get(url)
+                resp = await fetch_guarded(client, url)
                 resp.raise_for_status()
+        except BlockedURLError as exc:
+            return [{"type": "text", "text": f"Refused to fetch {url}: {exc}"}]
         except httpx.HTTPStatusError as exc:
             return [{"type": "text", "text": f"HTTP error {exc.response.status_code} fetching {url}"}]
         except Exception as exc:
diff --git a/backend/tests/test_url_guard.py b/backend/tests/test_url_guard.py
new file mode 100644
index 000000000..8cc59ca22
--- /dev/null
+++ b/backend/tests/test_url_guard.py
@@ -0,0 +1,101 @@
+"""SSRF guard: WebFetch must refuse non-http(s) schemes and any host that
+resolves to a private/loopback/link-local/cloud-metadata address, including
+after a redirect 30x-pivots toward the LAN."""
+import asyncio
+
+import pytest
+
+from backend.apps.agents.tools.url_guard import (
+    BlockedURLError,
+    fetch_guarded,
+    validate_fetch_url,
+)
+
+
+# ---------------- validate_fetch_url: schemes ----------------
+
+@pytest.mark.parametrize("url", [
+    "file:///etc/passwd",
+    "ftp://internal/secret",
+    "gopher://169.254.169.254/",
+    "data:text/html,<script>",
+    "//169.254.169.254/",
+])
+def test_rejects_non_http_schemes(url):
+    with pytest.raises(BlockedURLError):
+        validate_fetch_url(url)
+
+
+# ---------------- validate_fetch_url: IP literals ----------------
+
+@pytest.mark.parametrize("url", [
+    "http://169.254.169.254/latest/meta-data/",   # AWS/GCP metadata
+    "http://127.0.0.1/",                            # loopback
+    "http://localhost/",                            # loopback by name (resolves)
+    "http://10.0.0.5/",                             # private
+    "http://192.168.1.1/",                          # private
+    "http://172.16.0.1/",                           # private
+    "http://[::1]/",                                # ipv6 loopback
+    "http://[fd00::1]/",                            # ipv6 unique-local
+    "http://0.0.0.0/",                              # unspecified
+])
+def test_rejects_internal_ip_literals(url):
+    with pytest.raises(BlockedURLError):
+        validate_fetch_url(url)
+
+
+def test_rejects_url_with_no_host():
+    with pytest.raises(BlockedURLError):
+        validate_fetch_url("http:///nohost")
+
+
+def test_allows_public_ip_literal():
+    validate_fetch_url("https://8.8.8.8/")  # public, should not raise
+
+
+# ---------------- fetch_guarded: redirect pivot ----------------
+
+class _Resp:
+    def __init__(self, location=None):
+        self.is_redirect = location is not None
+        self.headers = {"location": location} if location else {}
+        self.next_request = None
+        if location:
+            class _Req:
+                url = location
+            self.next_request = _Req()
+
+
+class _FakeClient:
+    """Records every URL it was asked to GET so we can assert the guard
+    blocks before the request, not after."""
+    def __init__(self, script):
+        self._script = script
+        self.requested = []
+
+    async def get(self, url):
+        self.requested.append(url)
+        return self._script.pop(0)
+
+
+def test_redirect_to_internal_is_blocked_before_request():
+    # public URL 30x to metadata; the second hop must be validated and blocked,
+    # and we must never GET the internal address.
+    client = _FakeClient([_Resp(location="http://169.254.169.254/latest/")])
+    with pytest.raises(BlockedURLError):
+        asyncio.run(fetch_guarded(client, "https://example.com/start"))
+    assert client.requested == ["https://example.com/start"]
+
+
+def test_redirect_loop_caps_out():
+    script = [_Resp(location="https://example.com/again") for _ in range(20)]
+    client = _FakeClient(script)
+    with pytest.raises(BlockedURLError):
+        asyncio.run(fetch_guarded(client, "https://example.com/start"))
+
+
+def test_non_redirect_response_returned():
+    final = _Resp()
+    client = _FakeClient([final])
+    out = asyncio.run(fetch_guarded(client, "https://example.com/page"))
+    assert out is final