Files
marqdo-cli/docs/adr/0021-linux-strict-isolation.md
T
cflmy b6fc18233d feat(sandbox): implement probe-gated Linux strict isolation (ADR-0021)
Strict mode now re-executes the CLI inside a Bubblewrap supervisor with
user/mount/PID/net/UTS/IPC namespaces, no-new-privs, all capabilities dropped,
a read-only /project, private /proc and /dev, rlimits, output capture and
wall-clock process-group cleanup. The child verifies PID 1, NoNewPrivs, empty
CapEff, an unshared uid map and loopback-only networking before project code.

- PlatformCapabilities::detect runs a real bwrap probe; no probe, no strict
- strict L1 supports pure Marqdo + Math; Foreign/ABI/worker/kernel/grants/trace
  are rejected with exit 2 before sandboxing
- forged MARQDO_STRICT_ACTIVE is rejected; missing bwrap fails closed
- add adversarial helper, 7-test escape/resource matrix, seeded security lane
  script and self-hosted security workflow
- executed G-C lane as uid 1000 in a disposable Ubuntu 26.04 VM: 7/7 PASS
  plus positive CLI and refusal matrix; evidence in docs/evidence
- ADR-0021, spec/security.md, STATUS/COMPATIBILITY/README/CHANGELOG updated;
  ADR-0019 remains for non-Linux/fail-closed rules
2026-11-20 15:01:49 +08:00

4.7 KiB

ADR-0021: Linux strict isolation via Bubblewrap, namespaces and rlimits

Status: Accepted Date: 2026-11-20 Relates to: ADR-0004, ADR-0019, spec/security.md, stable-release plan §10 Security evidence: docs/evidence/2026-11-20-strict-gc.md

Context

ADR-0019 kept strict mode Unavailable everywhere because no adapter had been executed against the full escape/resource matrix, and the first development host could not create any namespace. A disposable Ubuntu 26.04 VM with a real kernel, writable cgroup2, passwordless sudo for test setup and the packaged Bubblewrap AppArmor profile became available. It can create user, mount, PID, network, UTS and IPC namespaces as an ordinary user through bwrap.

Decision

1. Dynamic, probe-gated availability

PlatformCapabilities::detect() on Linux runs a real Bubblewrap smoke test (--unshare-all --as-pid-1, read-only runtime bind, private /proc and /dev). Strict is Available { mechanism: "bubblewrap ..." } only when that probe succeeds; otherwise it stays Unavailable with the real reason. There is still no fallback to project.

2. L1 scope: pure language and native Math only

Strict accepts run/call of documents using the interpreter, console and native Math. It rejects --allow-read, --allow-write, --allow-foreign, --allow-plugin, --plugin, --trust-native-plugin, --kernel and --trace-file with exit 2 before any sandbox starts. Native ABI, worker plugins and Foreign kernels are therefore not reachable in strict mode in this release; L2 inheritance is future work.

3. Construction

The trusted supervisor is this process; the restricted child is the same marqdo executable re-executed under Bubblewrap:

  • --unshare-all --as-pid-1 --die-with-parent --new-session — user, mount, PID, network, UTS, IPC and cgroup namespaces; the child is PID 1;
  • read-only binds of /usr, /bin, /sbin, /lib, /lib64 and /etc/ld.so.cache, and of the project tree at /project;
  • a read-only bind of the program at /run/marqdo-program; no host /root, $HOME, /etc or other project directories are visible;
  • private /proc, minimal /dev, tmpfs /tmp;
  • --clearenv, no_new_privs, --cap-drop ALL;
  • hard rlimits: CPU, address space (512 MiB), file size, open files (64), processes (64), stack (8 MiB), core (0);
  • a supervisor deadline that kills the process group, and bounded capture of stdout/stderr.

The document argument is rewritten to /project/...; only the document token changes, all other arguments are preserved.

4. Child self-verification

The child accepts MARQDO_STRICT_ACTIVE=1 only if it can verify all of: getpid() == 1, NoNewPrivs: 1, empty CapEff, a non-initial uid map, and no non-loopback network interface. A forged marker outside the sandbox fails before project code runs.

5. Threat model and non-goals

This protects against untrusted Marqdo documents that use the language, interpreter, Math, console, or attempt direct syscalls. It does not sandbox native plugins (rejected), Foreign kernels (rejected), or a malicious kernel. Bubblewrap is the construction mechanism, not a certificate; the runtime probe and self-verification are part of the contract.

Known limits: rlimits are used instead of cgroup-v2 accounting (the VM has writable cgroups, but the released adapter does not yet claim cgroup enforcement); filesystem TOCTOU is limited by read-only mounts but not proven against a racing host user; the project tree is exposed read-only in full; network is isolated by a network namespace rather than a seccomp allowlist. No worker/Foreign L2 claim is made.

Evidence

Executed in a disposable Ubuntu 26.04 VM (kernel 7.0.0, container: none, Bubblewrap 0.11.1, AppArmor bwrap-userns-restrict profile, ordinary user dev):

  • 7/7 adversarial tests: identity, host reads, writes, /proc/privileges, network, fd/process limits, memory/CPU/output limits;
  • CLI: run, call, absolute paths, multi-module, default index.mq.md and --json all succeed under strict; Foreign/ABI/plugin/kernel/trace flags are rejected with exit 2;
  • forged MARQDO_STRICT_ACTIVE=1 outside the sandbox exits 1;
  • removing /usr/bin/bwrap makes strict Unavailable and the run exits 1 with no project execution.

Consequences

  • Linux strict is available only where the probe succeeds; this container still reports unavailable.
  • worker_processes remains true for crash isolation, but that is not part of the strict claim.
  • A future L2 ADR must extend the same supervisor to worker/Foreign processes, with its own escape tests, before those may run under strict.
  • A remote self-hosted security-lane URL is still required before a release calls G-C CI-verified; the local VM evidence is recorded separately.