runectl executes model-authored shell commands against CTF challenge material. Both are
untrusted input. This file states the containment boundary, its limits, and the reporting
process.
Open a private security advisory on this repository. Do not open a public issue for anything that would give an attacker a working attack before a fix exists.
Email alternative: hello@idalia.dev, the address published at
idalia.dev/.well-known/security.txt
(RFC 9116). It is not a dedicated security mailbox and reaches the same maintainer.
Acknowledgement target is a few days. This is a solo project; that is an intent, not an SLA.
In scope: sandbox escape or containment bypass; secret leakage into traces or logs; any path that executes model-authored code outside the container; any path that lets challenge material reach the host filesystem.
Out of scope: the agent writing a destructive command inside its own container (that
is the container's purpose); the model producing a wrong flag (see
docs/ARCHITECTURE.md D15); the cost of authorized API calls.
Purpose of the sandbox. Containment for two things expected to misbehave: commands an LLM wrote, and files a CTF author wrote. A wrong command destroys a disposable container rather than a working tree, and a challenge binary executes somewhere other than the host filesystem.
Limits of the sandbox. It is Docker. Docker is not a hardened boundary against an
attacker with a kernel exploit. runectl run is not appropriate for material believed to be
hostile — malware samples, untrusted binaries, challenges from an unvetted source. Those
require a virtual machine between the container and the host.
Applied per run (src/runectl/sandbox/docker.py):
| Control | Detail |
|---|---|
| One container per run | Named for the run id, removed on stop; no state carries between runs |
| Non-root, verified | USER ctf in the image; the run refuses to start if id -u returns 0. The check exists because arena ensure --from-file/--from-registry allows an arbitrary image to become the arena |
no-new-privileges |
Set as a security_opt |
| Unprivileged | privileged=False; no added capabilities, no device or socket mounts |
| Resource caps | 2 GB memory, 2 CPUs, 512 pids |
| Network off where possible | crypto and misc run network=none. web, pwn, rev, forensics, osint and network default to bridge (2026-09-09) for remote targets and live lookups. Set per category, overridable with --network |
| No host mounts | Challenge files are copied in; the host filesystem is never bind-mounted |
Not applied: no seccomp or AppArmor profile beyond Docker's default; no user-namespace
remapping; no egress filtering when the network is on — a web run can reach anything the
host can reach. Any of these mattering means running the whole thing inside a VM.
- Key precedence:
--api-key> environment > OS keyring >~/.config/runectl/keys.json. The keyring is the intended route (runectl keys set). The JSON fallback is created0600atomically, never written wide and then narrowed. - Keys never enter the sandbox. The container receives
TERMand nothing else. The agent cannot read credentials that are not present. - Traces are redacted on write. Every event passes a redaction hook before reaching disk, matching common provider key shapes. It is deliberately over-eager: a false positive is free, a leaked key is not. It is a backstop that matches known patterns, not a guarantee.
A trace records the challenge description, every command run, and every model response. That is its purpose, and it means a trace can contain challenge material that is not redistributable, plus any secret that appeared in command output without matching a redaction pattern. Review a trace before attaching it to a public issue.