Skip to content

Sandboxing ​

OrbySystems never gives an LLM, or code an LLM writes or hands off to, the run of your machine. Anything that executes Python or other arbitrary code runs inside a sandbox — limited to exactly what it needs to see on disk and reach over the network to do its job.

The general sandbox ​

The model's Python tool — used when it writes and runs a script, e.g. for analysis — runs inside a sandbox that:

  • can read and write only its own dedicated workspace, never your personal files or the rest of your home directory
  • has no network access by default; only hosts you explicitly add to the allowlist in Settings → Agent → Sandbox are reachable, for cases like installing a Python library a generated script depends on

The isolated parser sandbox ​

Bank statement and CSV parsers — both the ones bundled with OrbySystems and any you've dropped in yourself — run in a second, separate sandbox dedicated just to parsing, whether that's a real import, previewing a parser you're building, or running its tests. This sandbox denies network access entirely, with no allowlist at all: a parser's only job is turning a document already on your disk into structured data, so it never has a legitimate reason to reach the network — meaning even a malicious or buggy parser can't exfiltrate anything from the statement it's reading.

Privacy-centric AI agent