Files
openworker/coworker/web/guard.py
T
Mr-Neutr0n ff86735cf0 security: block loopback/private/metadata addresses in model-supplied URL fetches
web_fetch and browser_read_url take a URL straight from the model. The model's
input is untrusted by design - both tools' own descriptions call fetched
content "data to evaluate, not instructions" - and web_fetch is
requires_approval=False, so nothing prompts the user before the request goes
out.

Neither validated the address. Verified against a scratch server on loopback:

    web_fetch("http://127.0.0.1:9931/")
    -> {"text": "Directory listing for /\n.git/\n.github/..."}

No prompt, no error. The same call reaches http://169.254.169.254/ for cloud
metadata when OpenWorker runs on a VM, an Ollama instance on :11434, or any
service on the user's LAN. It cannot reach OpenWorker's own sidecar, which
requires COWORKER_API_TOKEN.

Adds coworker/web/guard.py: resolve the host and refuse when any answer lands
in loopback, private, link-local (which covers the metadata endpoint),
multicast or reserved space. Checking every resolved address means a name with
one public and one private A record is refused rather than raced.

Redirects are the usual bypass, so follow_redirects is off and the chain is
walked here with each hop checked before it is requested. _request grows an
opt-in check_addresses flag used only by browser_read_url; the hardcoded vendor
endpoints the rest of the connectors call skip the guard and its DNS lookup.

Not covered, and stated in the module docstring: DNS rebinding. The name is
resolved by the guard and again by the client when it connects, so a near-zero
TTL record can change in between. Closing that needs connection-level IP
pinning. The hop check is the cheap 90%.

Tests: tests/test_url_address_guard.py - literals, IPv4-mapped IPv6 loopback,
names resolving into private space, split-horizon answers, non-http schemes,
redirect into loopback proven not to be requested, and a bounded redirect
loop.
2026-07-29 02:22:10 +05:30

110 lines
4.0 KiB
Python

"""Address guard for URLs the model chooses.
`web_fetch` and `browser_read_url` take a URL straight from the model, and the model's
input is untrusted by design — it reads web pages, email and Slack messages, all of which
are documented as "data, not instructions". A page that talks the agent into fetching
`http://169.254.169.254/` or `http://127.0.0.1:11434/` turns a read-only research tool into
a probe of the machine's own network position, and `web_fetch` is `requires_approval=False`,
so no prompt ever appears.
This blocks the ranges that are only reachable *because* OpenWorker runs on the user's
machine: loopback, RFC1918 and other private space, link-local (which covers the cloud
metadata endpoint at 169.254.169.254), and the reserved/multicast blocks.
Every hop is checked, not just the first: `follow_redirects=True` otherwise lets a public
URL 302 straight to loopback, which is the standard way this filter is bypassed.
Not covered: DNS rebinding. The name is resolved here and resolved again by the client when
it connects, so a record with a ~0 TTL can change between the two. Closing that needs
connection-level IP pinning; the hop check is the cheap 90% and is stated as such.
"""
from __future__ import annotations
import ipaddress
import socket
from typing import Optional
from urllib.parse import urlsplit
MAX_REDIRECTS = 5
def _blocked_reason(ip: ipaddress._BaseAddress) -> Optional[str]:
if ip.is_loopback:
return "loopback"
if ip.is_link_local:
return "link-local (includes the cloud metadata endpoint)"
if ip.is_private:
return "a private network"
if ip.is_multicast:
return "multicast"
if ip.is_reserved or ip.is_unspecified:
return "a reserved range"
return None
def check_url(url: str) -> Optional[str]:
"""None if the URL may be fetched, else a human-readable refusal reason.
Resolves the host and rejects when *any* answer lands in a blocked range, so a name
with both a public and a private A record cannot be used to slip through.
"""
parts = urlsplit(url)
if parts.scheme not in ("http", "https"):
return "url must start with http:// or https://"
host = parts.hostname
if not host:
return "url has no host"
# A literal address needs no lookup.
try:
literal = ipaddress.ip_address(host)
except ValueError:
literal = None
if literal is not None:
reason = _blocked_reason(literal)
return f"refusing to fetch {host}: {reason}" if reason else None
try:
infos = socket.getaddrinfo(host, parts.port or (443 if parts.scheme == "https" else 80),
proto=socket.IPPROTO_TCP)
except OSError as exc:
return f"could not resolve {host}: {exc}"
for info in infos:
raw = info[4][0]
try:
ip = ipaddress.ip_address(raw)
except ValueError:
continue
# ::ffff:127.0.0.1 and friends must be judged as the v4 address they carry.
mapped = getattr(ip, "ipv4_mapped", None)
if mapped is not None:
ip = mapped
reason = _blocked_reason(ip)
if reason:
return f"refusing to fetch {host} ({ip}): {reason}"
return None
def get_checked(client, url: str, *, max_redirects: int = MAX_REDIRECTS):
"""GET `url`, validating the address before every hop.
`client` must be built with `follow_redirects=False`; redirects are walked here so each
Location is checked. Returns the final response. Raises `PermissionError` when a hop is
refused, `RuntimeError` when the redirect budget is exhausted.
"""
seen = url
for _ in range(max_redirects + 1):
reason = check_url(seen)
if reason:
raise PermissionError(reason)
resp = client.get(seen)
if resp.status_code not in (301, 302, 303, 307, 308):
return resp
location = resp.headers.get("location")
if not location:
return resp
seen = str(resp.url.join(location))
raise RuntimeError(f"too many redirects (>{max_redirects})")