The container intentionally restricts certain operations:
| Resource | Access | Purpose |
|---|---|---|
| Network | proxy-filtered (Vertex AI, PyPI, GitHub, agent-specific) | Prevents data exfiltration |
| Current directory | read-write | Working files |
~/.config/gcloud |
read-only | Vertex AI auth |
| Agent config | copied in, not mounted | Prevents host config poisoning |
~/.gitconfig |
read-only | Git identity |
| SSH keys | not mounted | Prevents git push via SSH |
| GitHub CLI config | not mounted (uses /tmp/gh-config) | Prevents cached host credentials |
GH_TOKEN (host) |
never propagated | Use PAUDE_GITHUB_TOKEN or --github-token on start/connect |
| Git credentials | not mounted | Prevents HTTPS git push |
These exfiltration paths have been tested and confirmed blocked:
| Attack Vector | Status | How |
|---|---|---|
| HTTP/HTTPS exfiltration | Blocked | Internal network has no external DNS; proxy allowlists only approved domains |
| Git push via SSH | Blocked | No ~/.ssh mounted; DNS resolution fails anyway |
| Git push via HTTPS | Blocked | No credential helpers; no stored credentials; DNS blocked |
| GitHub CLI write ops | Relies on token scope — use a read-only fine-grained PAT | Use read-only PAT via PAUDE_GITHUB_TOKEN; host GH_TOKEN never propagated |
| Modify cloud credentials | Blocked | gcloud directory mounted read-only |
| Escape container | Blocked | Non-root user; standard Podman isolation |
# SAFE: Network filtered, cannot exfiltrate data
paude create --yolo
# DANGEROUS: Full network access, can send files anywhere
paude create --yolo --allowed-domains allThe --yolo flag enables autonomous execution (no confirmation prompts). This is safe when network filtering is active because the agent cannot exfiltrate files or secrets even if it reads them.
Do not combine --yolo with --allowed-domains all unless you fully trust the task.
The container has full read-write access to your working directory. Your protection is git itself. Push important work to a remote before running in autonomous mode:
git push origin mainIf something goes wrong, recovery is a clone away.
These risks are accepted by design:
- Workspace destruction: The agent can delete files including
.git. Mitigation: push to remote before autonomous sessions. - Secrets readable:
.envfiles in workspace are readable. Mitigation: network filtering prevents exfiltration; don't use--allowed-domains allwith sensitive workspaces. - No audit logging: Commands executed aren't logged. This is a forensics gap, not a security breach vector.
These properties are ignored for security reasons:
mounts- paude controls mountsrunArgs- paude controls run argumentsprivileged- never allowedcapAdd- never allowedforwardPorts- paude controls networkingremoteUser- paude controls user