run timeout sends kill(-1, SIGTERM) — kills ALL processes on the host, not just the sandbox

[xulis]

Description

When a run command exceeds resources.timeout_ms, finSAFE executes kill(-1, SIGTERM) which sends SIGTERM to every process on the host (all PIDs except itself), instead of only terminating the sandboxed process tree (e.g. via cgroup).

Environment

• finSAFE 0.9.34 (personal CLI, install.sh)
• Linux 6.8.0-124-generic, x86_64, Ubuntu 24.04
• cgroup v2 (finsafe.slice), bwrap available

Steps to reproduce

1. Create a policy with a short timeout:

resources:
  memory_max: "64M"
  pids_max: "32"
  cpu_max: "50000 100000"
  timeout_ms: 5000

2. Run a command that exceeds the timeout:

finsafe --policy timeout.yaml run -- sleep 10

3. Observe: the SSH session running the command is disconnected, and the parent shell is killed.

Evidence (strace)

13822 kill(-1, SIGTERM)                 = 0
13821 +++ killed by SIGTERM +++
13816 +++ killed by SIGTERM +++
13818 +++ killed by SIGTERM +++
13817 +++ killed by SIGTERM +++
13804 +++ killed by SIGTERM +++

The strace shows kill(-1, SIGTERM) — the parent test script (PID 13791), the SSH session, and unrelated processes all received SIGTERM.

Expected behavior

Timeout should terminate only the sandboxed process (and its descendants), e.g. via the cgroup cgroup.kill or by killing the sandbox PID. The host shell/SSH session must be unaffected.

Impact

• Running finsafe run with a timeout in a CI job or over SSH kills the whole session (and any other processes on the machine).
• Any process on the host can be SIGTERMed — in the worst case, a long-running server process gets terminated. (SIGKILL would be even worse.)

[Geeksfino]

Automated fix: A pull request is open for this issue.

• PR: Geeksfino/finsafe#205
• Status: In review — merge and public release are handled by a separate ship step.
• What changed: Timeout cleanup no longer calls kill(-1, SIGTERM). When your policy creates a per-run cgroup, timeout uses cgroup.kill so only sandbox processes are stopped; otherwise finSAFE signals only the sandbox child (not your SSH/CI shell).
• You will get another comment here when a public release is published.

[Geeksfino]

Automated dispatch: Cloud agent started for Geeksfino/finsafe.

• Agent: bc-2f83f242-a110-4e4e-b6df-a44b51453a91
• Run: run-929c593b-404f-4d86-8d8f-fdc348dba5f5

A PR will be opened when the agent finishes. Fixes will reference this issue.

[Geeksfino]

Resolved — fix shipped in a public release.

• Version: v0.9.35
• Release: https://github.com/finogeeks/finsafe/releases/tag/v0.9.35
• Merge: Geeksfino/finsafe#205
• Summary: fix(linux): scope run timeout kill to sandbox only

Install or upgrade: see the release assets on the link above.

[xulis]

Verified fixed in 0.9.35

Updated to 0.9.35 and re-ran the reproducer:

finsafe --policy timeout.yaml run -- sleep 10   # timeout_ms: 3000

Result: termination_reason=timeout_killed exit_code=None timed_out=true — the SSH session stays alive, the parent script survives.

strace confirms the fix:

21156 kill(-21168, SIGTERM) ... = 0     # negative PID = sandbox process group only
21173 +++ killed by SIGKILL +++
21172 +++ killed by SIGKILL +++
21168 +++ killed by SIGKILL +++

kill(-1, SIGTERM) is gone — timeout now signals only the sandbox process group (negative PID), not all host processes. The parent shell prints its final line after the timeout kill.

Thank you!