READMEs are untrusted code. readme2demo executes them, so isolation is a core design concern, not an afterthought.
The agent and verify stages run inside hardened Docker containers (Sandbox.start()):
--cap-drop ALLand--security-opt no-new-privileges- non-root user (
USER demoin the base image) - memory, CPU, and PID limits
- the repo is mounted read-only; the agent works on a writable copy
The verify stage then replays the distilled commands.sh in a separate,
fresh container with no model credentials and no state carried over from the
agent's run.
For a claims-vs-roadmap table (enforced today vs planned v0.8 vs known
limitations, each with code pointers), see the root
SECURITY.md. That table is the source of truth; this page
is the short summary.
-
Render container is not fully hardened like agent/verify.
render.pyruns its owndocker runwith memory/CPU/network limits only (no cap-drop / no-new-privileges / pids-limit). See the root SECURITY.md table. -
API key in the sandbox (MVP): the model credential currently enters the sandbox (and the
docker runargv). Use a dedicated, low-limit key. A host-side, key-injecting egress proxy is planned so the credential never crosses the boundary (#64; argv exposure tracked in #51). -
--allow-docker-socketis off by default. Enabling it mounts the host Docker socket into the sandbox so demos that manage containers work — but the socket pierces isolation. Only enable it for repositories you trust. -
Network egress is open bridge networking today, not allowlisted. The sandbox defaults to Docker
--network bridgeso package installs and git clones work. Domain allowlisting is planned for v0.8 (#64). Until then, treat outbound network from a compromised agent as possible.
Please report vulnerabilities privately, not in public issues. See
SECURITY.md
in the repository for the full threat model and the private reporting channel
(GitHub Security Advisories).