Commit Graph
2 Commits
Author SHA1 Message Date
23c37b9c98 feat(container): restrict sockets and process tracing in Compose (#160)
* Harden sandbox seccomp: drop ptrace and tighten socket domain allowlist

Two fixes from a security audit of this repo:

- Remove process_vm_readv/process_vm_writev/ptrace from the allowlist.
  All sandboxed processes share the same uid, so ptrace let one process
  inspect/modify another's memory (e.g. read a secret held by a sibling
  process) inside the sandbox -- inter-process isolation was effectively
  void despite the seccomp profile's intent.

- Replace the 3 socket-domain rules with an explicit allowlist of
  AF_UNIX(1)/AF_INET(2)/AF_INET6(10). The previous rules (domain<38,
  domain==39, domain>40) only actually blocked AF_ALG(38) and
  AF_VSOCK(40) -- AF_NETLINK(16) and AF_PACKET(17) were allowed, along
  with everything else outside that narrow gap, making the egress
  control close to a no-op.

Verified with a real Docker build + run on a separate host, under the
exact production security contract (--cap-drop ALL --security-opt
no-new-privileges + this seccomp profile):
- DNS resolution (getaddrinfo) and `git clone` over HTTPS still work
  (AF_INET/AF_INET6 unaffected) -- confirms this doesn't break the
  scanner's own legitimate network use.
- AF_NETLINK and AF_PACKET socket() calls now fail with EPERM inside
  the container (previously succeeded) -- confirms the fix is real, not
  cosmetic.

Local commit only -- this is an audit clone of openai/codex-security,
not a fork with push access.

* Fail closed if not launched with --cap-drop ALL --security-opt no-new-privileges

docker/codex-security-seccomp.json unconditionally allows mount/pivot_root/
unshare/setns/clone for Codex's own unprivileged Bubblewrap sandboxing --
those syscalls can't be removed without breaking that feature entirely.
Their safety previously depended entirely on a comment in the seccomp
profile asserting the caller runs with --cap-drop ALL and
--security-opt no-new-privileges (true for compose.yaml, not enforced for
a bare `docker run`).

Adds a runtime self-check at the top of the entrypoint: read
/proc/self/status, refuse to start (exit 3, clear message) unless CapEff
is all-zero and NoNewPrivs is 1. Converts an undocumented-until-you-read-
the-seccomp-comment assumption into a loud, immediate failure.

Verified with a real Docker build + run on a separate host, 4 launch
scenarios:
- --cap-drop ALL --security-opt no-new-privileges (contract met): starts
  normally.
- No flags, default non-root user (10001 from the Dockerfile): blocked by
  the NoNewPrivs check.
- --security-opt no-new-privileges only, no --cap-drop: starts -- correct,
  not a false block, since a non-root process already has an empty
  effective capability set regardless of --cap-drop.
- --user root, no flags: blocked by the CapEff check, message shows the
  actual non-empty capability bitmask (00000000a80425fb).

Local commit only -- this is an audit clone of openai/codex-security, not
a fork with push access.

* Address Codex review: CI hardening flags, AF_NETLINK for bwrap, drop socketcall

Three fixes for the automated review findings on this PR:

1. .github/workflows/container-ci.yml: the bare `docker run` smoke-test
   invocations (Verify bundled scanner, Reject interactive repository
   discovery) never passed --cap-drop ALL --security-opt no-new-privileges,
   so entrypoint.sh's new fail-closed check aborted them with exit 3 before
   --version/help/discovery validation could run. Added the flags to match
   the production contract these tests are supposed to exercise.

2. docker/codex-security-seccomp.json: the AF_UNIX/AF_INET/AF_INET6-only
   socket() allowlist rejected AF_NETLINK, which Bubblewrap needs (a
   NETLINK_ROUTE socket) to run RTM_NEWLINK/RTM_NEWADDR when configuring
   the loopback interface inside its own isolated network namespace during
   codex sandbox startup. The CI workflow already silently tolerated this
   by retrying with --enable use_legacy_landlock; added an explicit
   AF_NETLINK(16) rule restricted to protocol NETLINK_ROUTE(0) so the
   normal Bubblewrap path succeeds without needing that fallback.

3. docker/codex-security-seccomp.json: "socketcall" was unconditionally
   allowed in the generic syscall list with no argument filtering. On the
   supported SCMP_ARCH_X86 subarchitecture, this legacy multiplexer lets a
   32-bit executable open sockets (including AF_NETLINK/AF_PACKET) via
   socketcall's own SYS_SOCKET dispatch, bypassing the domain-restricted
   socket() rules entirely -- seccomp can't filter socketcall by socket
   domain since that lives behind a pointer argument, not a plain integer,
   so the safe fix is to stop allowing it unconditionally. socketpair (only
   ever AF_UNIX by kernel definition) stays allowed.

Verified with a real Docker build + container runs on a separate host
(not the CI runner, not the openai org):
- Bare `docker run` (no flags): still fails closed with exit 3, unchanged.
- With the added flags: --version, bulk-scan --help, info --json all
  exit 0; the discovery-rejection test still returns exit 2 with the
  expected message.
- A/B on the sandbox seccomp: the pre-fix profile reproduces the exact
  error from the review ("bwrap: loopback: Failed to create NETLINK_ROUTE
  socket: Operation not permitted"); the fixed profile runs the same
  `codex sandbox /usr/bin/true` command to completion (exit 0) without
  the use_legacy_landlock fallback.
- Normal HTTPS egress (AF_INET, via Node's https module) still works
  under the fixed profile, confirming the socketcall removal doesn't
  regress git-clone/API access on x86_64.

---------

Co-authored-by: mldangelo-oai <mdangelo@openai.com>
Co-authored-by: mldangelo-oai <269034524+mldangelo-oai@users.noreply.github.com>
2026-11-04 12:06:09 -08:00
ianw-oai 10e7a22461 chore(codex-security): sync public projection
- [codex-security] prevent untrusted pre-commit hook execut...
  (And 4 more changes)

GitOrigin-Timestamp=2026-07-27T10:45:08-07:00
GitOrigin-RevId: 23c8e48a18db9283b3fc1335bf40ea63402400c6
2026-07-27 18:00:36 +00:00