Trellis is only sanctioned for use on a dedicated machine that holds no private or valuable data. Please read this before running it.
Driving nondeterministic LLM agents to execute code carries inherent risk: their behavior is not fully predictable. Trellis has been developed and tuned for performance in hardened, isolated environments that a trusted operator controls — not yet for wide consumer use. Treat it as operator/research software and run it only where you accept these risks. It is provided AS IS, without warranty (see LICENSE).
Trellis drives external LLM agent CLIs (codex, claude, gemini) fully
autonomously, with their approval prompts disabled
(--dangerously-bypass-approvals-and-sandbox, --dangerously-skip-permissions,
--approval-mode=yolo). This is required for unattended multi-cycle operation,
but it means the agents execute commands with no human in the loop.
Each worker/reviewer burst runs inside a bubblewrap
(bwrap) mount sandbox that gives the agent a dedicated burst home and a
read-only view of the project repo (plus a few writable build directories),
drops capabilities (--cap-drop ALL), and unshares the pid/ipc/uts namespaces —
so the agent cannot casually read the rest of your home directory.
That sandbox is containment, not isolation — it is not a hard security boundary. Specifically:
- The burst runs as your own user. There is no privilege separation.
- It has full network access — agents must reach the provider APIs, git, and the mathlib cache, so egress cannot be cut.
- It can read the provider credentials it needs to authenticate: your
~/.codex,~/.claude, and~/.geminiare mounted read-only into the worker. - Unprivileged user namespaces have a non-trivial escape surface.
In short: a buggy, jailbroken, or prompt-injected agent can run arbitrary commands as you, read your LLM provider credentials, and use the open network to exfiltrate anything it can reach — the repo, the paper, and those credentials.
- Run only on a dedicated or throwaway machine / VM with no private or valuable data, no other accounts' secrets, and no production access.
- Use a dedicated provider account for the agent credentials — not your primary account — and revoke it when you are done.
- Keep
sandbox.enabled: true(the default). With the sandbox disabled, agents run directly on the host with no containment at all. - Treat the project repo and anything reachable from the host as exposed.
Please report security issues privately to wes@math.cmu.edu rather than opening a public issue.