fix: restrict docker commands for ai-worker user

Remove ai-worker from docker group and enforce sudo whitelist.

SECURITY: Being in the docker group gives unrestricted access to the
Docker daemon socket (/var/run/docker.sock), allowing any docker command:
docker exec, docker cp, docker run -v /:/host, docker commit, etc.

Changes:
- Remove extraGroups = ["docker"] from ai-worker user definition
- Add comprehensive sudo NOPASSWD whitelist for safe docker subcommands
  ALLOWED: ps, inspect, logs, images, info, version, stats, start, stop,
  restart, rm, rmi, wait, pull, build, run, compose, system,
  network ls, volume ls
  BLOCKED (implicitly): exec, cp, commit, diff, export, import, load,
  save, attach, push, tag, create, plugin, network create, volume create
- Update ai-worker-restricted.nix module to reflect new approach
- Update README-ai-worker.md with new security model and examples

All docker commands must now be prefixed with sudo.
The Hermes agent's host_run tool needs to be updated to prepend sudo.
This commit is contained in:
2026-05-20 20:34:19 -04:00
parent 36359de6aa
commit 9459839d74
3 changed files with 208 additions and 68 deletions

View File

@@ -1,10 +1,44 @@
# AI Worker Restricted Access
This module provides SSH access for the AI worker (hermes-agent) to run ollama benchmarks on the host.
This module provides SSH access for the AI worker (hermes-agent) to run docker commands on the host.
## Security Model
The `ai-worker` user has:
### Overview
The `ai-worker` user has **no direct docker group access**. All docker commands must go through `sudo`, and only specific subcommands are whitelisted:
- **Container lifecycle**: `docker ps`, `docker inspect`, `docker logs`, `docker images`, `docker info`, `docker version`, `docker stats`
- **Control**: `docker start`, `docker stop`, `docker restart`, `docker rm`, `docker rmi`, `docker wait`
- **Image management**: `docker pull`, `docker build`, `docker run`, `docker compose`
- **Disk cleanup**: `docker system`
- **Network/Volume**: `docker network ls`, `docker volume ls` (read-only)
### EXPLICITLY BLOCKED (not in sudo whitelist)
| Command | Risk | Result |
|---------|------|--------|
| `docker exec` | Execute arbitrary commands inside containers (FILE MODIFICATION) | Blocked by sudo |
| `docker cp` | Copy files between containers and host | Blocked by sudo |
| `docker commit` | Create images from running containers (data exfil) | Blocked by sudo |
| `docker diff` | Inspect filesystem changes | Blocked by sudo |
| `docker export` | Export container filesystem | Blocked by sudo |
| `docker import` | Import filesystem archives | Blocked by sudo |
| `docker load` | Load docker images | Blocked by sudo |
| `docker save` | Save docker images to tar | Blocked by sudo |
| `docker attach` | Interactive access to containers | Blocked by sudo |
| `docker push` | Push images to registries | Blocked by sudo |
| `docker tag` | Rename images | Blocked by sudo |
### Why This Approach?
Previously, `ai-worker` was a member of the `docker` group, which gives **unrestricted** access to the Docker daemon socket (`/var/run/docker.sock`). Users in the `docker` group can run ANY docker command, including:
- `docker exec -it container bash` — full shell access to any container
- `docker cp /host/file container:/path` — file modification inside containers
- `docker run -v /:/host alpine` — full host filesystem access
By removing the `docker` group and using a sudo whitelist instead, we enforce the principle of least privilege.
### Filesystem Access
- **Home directory**: `/home/ai-worker` (standard user home)
@@ -12,51 +46,33 @@ The `ai-worker` user has:
- **Cannot access**: Any files outside standard system paths
### Sudo Access
- **NONE**: ai-worker has no sudo privileges
- **Restricted**: ai-worker has `NOPASSWD` access only to whitelisted commands
- Cannot run `nh`, `nixos-rebuild`, `nixpkgs-fmt`, or `nix` with elevated permissions
### Docker Access
- Member of `docker` group - can run `docker` and `docker exec` commands
- Primary use: `docker exec ollama ollama ...` for benchmarking
- Can run `docker exec --privileged ollama rocm-smi ...` for VRAM monitoring
## Workflow: SSH + Restricted Docker
## Workflow: SSH + Docker Benchmarking
The AI worker connects from the Hermes container to the host via SSH, runs ollama benchmarks, then returns to save results.
### Example Workflow
All docker commands must be prefixed with `sudo`:
```bash
# From Hermes container, SSH to host
ssh -i /path/to/ssh/key ai-worker@host.docker.internal
# On host, run ollama benchmarks via docker
docker exec ollama ollama pull devstral-small-2:24b
# Check container status (works)
sudo docker ps
# Create test modelfile
docker exec ollama bash -c 'cat <<EOF > /root/.ollama/test.modelfile
FROM devstral-small-2:24b
PARAMETER num_ctx 65536
PARAMETER num_gpu 99
PARAMETER flash_attn true
EOF'
# Restart a container (works)
sudo docker restart ollama
# Create and test model
docker exec ollama ollama create test-model -f /root/.ollama/test.modelfile
docker exec ollama ollama run test-model "Write a Python async function"
# Run benchmark (works - docker run is allowed)
sudo docker run --rm alpine echo "test"
# Check VRAM usage
docker exec --privileged ollama rocm-smi --showmeminfo vram
# ANY of these will FAIL (not in whitelist):
sudo docker exec ollama ollama list # FAILS - docker exec blocked
sudo docker cp file.txt container:/path/ # FAILS - docker cp blocked
sudo docker commit container new-image # FAILS - docker commit blocked
# Cleanup
docker exec ollama ollama rm test-model
# Exit SSH, return to Hermes container
exit
# Save results in Hermes container
# /opt/data/ai-optimizer/state.json
# /opt/data/ai-optimizer/results.csv
# For ollama operations, use the HTTP API instead of docker exec:
curl http://ollama:11434/api/tags
```
## SSH Access
@@ -74,25 +90,30 @@ Check ai-worker permissions:
```bash
# On the host, as root or gortium:
sudo -u ai-worker sudo -l
# Should show: no sudo access
# Should show the whitelisted commands only (no docker exec/cp/commit)
# Check docker group membership
# Verify NOT in docker group
groups ai-worker
# Should show: ai-worker docker
# Should show: ai-worker (NO docker group)
```
## Troubleshooting
If ai-worker cannot run docker commands:
If docker commands fail:
```bash
# Check docker group membership
# Check sudo permissions
sudo -u ai-worker sudo -l | grep docker
# Verify group membership
groups ai-worker
# Verify ollama container is running
docker ps | grep ollama
# Test allowed command
sudo -u ai-worker sudo docker ps
# Test docker access
sudo -u ai-worker docker exec ollama ollama list
# Test blocked command (should fail)
sudo -u ai-worker sudo docker exec ollama ollama list
# Expected: "Sorry, user ai-worker is not allowed to execute"
```
If SSH connection fails: