| name | untrusted-execution |
| description | Security playbook for running third-party / untrusted code (research-paper repos, user-supplied scripts) inside DaxAlgo Terminal. Deny-by-default sandbox (Docker / WSL2 / VM), egress allowlist, CPU/RAM/pids/disk/wall-clock quotas, no host mounts of the canonical store or credentials, kill-the-process-tree, never exec in-process. Use whenever editing src/TradingTerminal.Infrastructure/Research/Sandbox/, ISandboxRunner, SandboxPolicy/SandboxQuota, the Paper Lab repro engine, or designing ANY feature that executes code the user did not write. This is a security subsystem, not a feature — read it before weakening any isolation. |
Untrusted Execution (the sandbox)
Running a research paper's real code means running arbitrary, untrusted code from arbitrary repos.
That is a remote-code-execution surface. The rule is simple: untrusted code runs only inside a
disposable container/VM, deny-by-default, and the C# side only orchestrates the sandbox over a
process boundary — it never execs paper code in-process.
Threat model
The reproduced code is hostile until proven otherwise. It may try to: exfiltrate the canonical store
or broker credentials, beacon out over the network, fork-bomb / exhaust RAM/disk, persist on the
host, or pivot to the broker sessions. Every mitigation below maps to one of these.
Non-negotiable controls (deny-by-default)
ISandboxRunner.RunAsync(ReproSpec, SandboxQuota, SandboxPolicy, IProgress<string>, ct) must enforce:
- Isolation: a container (Docker first) or VM/WSL2 — never the host process, never
Process.Start
of the paper's entrypoint directly. The runner spawns the container CLI, not the paper code.
- No host mounts of anything sensitive: NO bind-mount of the market-data store dir, the credential
store, the user profile, or the repo tree's parent. Only a scratch tmpfs/volume is writable.
- Read-only rootfs + tmpfs scratch; drop Linux capabilities (
--cap-drop ALL); --pids-limit;
non-privileged user; no --privileged, no docker socket mount.
- Network deny-by-default:
--network none. The SandboxPolicy egress allowlist is opt-in and
scoped to the paper's declared data deps only — never "allow all".
- Quotas from
SandboxQuota: --memory, --cpus, disk cap, and a wall-clock timeout. On
timeout/cancel, kill the entire process tree (clone LeanProcessRunner's kill-tree).
- Concurrency bound: a
SemaphoreSlim from SandboxOptions.MaxConcurrent — one runaway job must
not starve the box.
- One-way data flow out: results leave only as a declared artifact file in the scratch volume,
read back and validated by
IReproSignalBridge. The sandbox cannot write anywhere else.
Layer + build placement
SandboxPolicy / SandboxQuota / SandboxKind are Core/Research/ records/enums (inputs to the
runner, deny-by-default constructors).DockerSandboxRunner / Wsl2SandboxRunner / SandboxProcess / RepoFetcher live in
Infrastructure/Research/Sandbox/. Selected by SandboxOptions.Kind in AddPaperResearch.- No Python/native deps in the C# build — env-resolution + minimal-repro LLM logic live in the
Python sidecar (
tools/python-ml/); the sidecar itself binds 127.0.0.1 only and is also outside
the container's egress allowlist.
Cost + reliability
- The minimal repro runs first;
ReplicationCostEstimator extrapolates its wall-clock/RAM to the
full config (autoarxiv's "replication cost"). The full run is gated behind a budget the user
approves — never auto-escalate.
- Sandbox failures fold into a
Failed ReproResult (never throw across the seam), same never-throw
contract as IAiAnalystClient.
Verifier checklist (the deterministic invariants)
What NOT to do
- Don't bind-mount a host path "just to make it work" — that's the exfiltration hole.
- Don't default the egress allowlist to open, or skip it "for now".
- Don't run the repo on the host even once "to test" — there is no trusted first run.
- Don't let the sandbox write outside the scratch volume; don't read its stdout as the result —
read the declared, validated artifact.
- Don't throw across
ISandboxRunner — fold failures into ReproResult.Failed.
