Skip to content

Security

What the container trusts

Input Trust level Why
MULTICA_TOKEN Secret Full access to the workspaces the token can reach
AI_API_KEY Secret Billable access to the model endpoint
GH_TOKEN Secret Repository access, often write
Task instructions Untrusted content A prompt injection can ask the agent to exfiltrate anything the container can read

The agent runs arbitrary code by design. Treat the container as a sandbox for convenience, not as a security boundary against a malicious task. The token you give it is the blast radius.

Practical consequences:

  • Use a Multica token scoped to the workspaces the runtime needs.
  • Use a GitHub token scoped to the repositories the agent is allowed to touch.
  • Do not mount a host directory that contains unrelated credentials.
  • Do not run the autoscaler's Docker socket mount on a host where tasks are untrusted; the socket is root-equivalent.

Secrets handling

  • No secret is baked into an image layer. Everything arrives as an environment variable or an orchestrator secret at runtime.
  • The OpenCode config contains the API key and is written at start with mode 0600 under the runner's home.
  • The banner never prints a token, key, or full router URL (only scheme + host).
  • .env is git-ignored; .env.example contains empty placeholders.
  • ECS task definitions reference Secrets Manager ARNs, not values.

If a key leaks: rotate it at the provider, then restart the runtime. The old key is not recoverable from the image, but it may be in logs if you printed it yourself.

Non-root

Everything runs as runner (uid/gid 1000):

  • the Multica daemon, both agent CLIs, and every task;
  • Xvfb and the Penpot MCP server in the design image;
  • MariaDB, PostgreSQL and Redis in the PHP image.

There is no sudo, no setuid helper, and no runtime privilege drop. This keeps the image compatible with Fargate's seccomp profile and with Kubernetes runAsNonRoot.

Network

  • No inbound port. The runtime connects out; it does not listen for tasks.
  • Egress should be restricted to what the runtime needs: Multica, the AI router, your Git host, and the package registries your tasks use. A coding agent needs broad egress in practice; an allowlist is a policy decision, not a default.
  • The design image binds MCP and plugin servers to 127.0.0.1 by default. Change the host variables only if you deliberately want them reachable.

Supply chain

  • Base OS, Node, and every tool are pinned in versions.env.
  • Downloads are checksum-verified where upstream publishes a checksum.
  • Images are built in GitHub Actions, not on a laptop, and published to GHCR with provenance and SBOM attestations.
  • Dependencies are updated by Dependabot PRs; a human reviews the diff and the resulting image.

For high-assurance deployments:

docker pull ghcr.io/proreka/multica-runtime@sha256:<digest>

Pin the digest, not the tag. Tags are mutable; digests are not.

What is deliberately not done

  • No auto-update of agent CLIs. An in-place update would change behaviour mid-task and bypass review.
  • No Docker socket except the Swarm autoscaler. The ECS adapter uses IAM.
  • No privileged mode, no host PID/network, no --cap-add.
  • No secrets in build args. Build args are visible in image history.

Reporting a vulnerability

Please do not open a public issue for a security problem. See SECURITY.md for the private reporting channel and response expectations.