Skip to content
Docs

Security Posture

Lifeboat runs on infrastructure you own, so its security model assumes you control the network and the host, and it secures what it is responsible for.

Access control

Section titled “Access control”

Four roles, least-privilege by default, covered in Users and roles. Applications get API keys scoped to inference, revocable individually.

No default credentials ship. A fresh deployment asks you to create the first administrator; if no password is supplied the container generates one per install.

Secrets at rest

Section titled “Secrets at rest”
  • Hugging Face tokens are encrypted in the credential vault.
  • API keys are stored so that an administrator can reveal one deliberately, and every reveal is audited.
  • Model encryption keys, where confidential computing is in use, are sealed and released only while the host attests. See Encrypted weights.

Fleet authentication

Section titled “Fleet authentication”

Control plane to node agent is authenticated with a shared secret. A wrong or missing secret is rejected with the node named, rather than producing a node that silently never appears.

Container hardening

Section titled “Container hardening”

The container runs as a non-root user with capabilities dropped, and the application directory is owned by that user. The Helm chart sets the user id explicitly for the same reason.

One documented exception: AMD’s own container guidance requires a relaxed seccomp profile for ROCm’s userspace device calls. That is their requirement rather than a Lifeboat choice, and it is the one place the AMD posture is weaker than NVIDIA’s.

Network

Section titled “Network”

Lifeboat does not assume it is on a trusted network but does not replace one. Put TLS in front, or give it a certificate and key. Model server ports do not need to be exposed beyond the host — only the control-plane port does.

What is deliberately out of scope

Section titled “What is deliberately out of scope”

Stated plainly, because an evaluator should not have to discover it:

  • Data-loss prevention, prompt guardrails, per-team budgets and cross-provider governance. Those belong in a governance layer in front of inference; Lifeboat does not duplicate one.
  • A SIEM. The audit log is complete and exportable; correlation belongs in your existing platform.
  • Prompt and completion logging. Lifeboat does not retain request content, which is also why none of it appears in telemetry.

What to check when evaluating

Section titled “What to check when evaluating”
  1. That the data volume is persisted and backed up — it carries the audit log, keys and the licence binding.
  2. That the control-plane port is fronted by your own TLS and access control.
  3. That model server ports are not exposed — a client hitting a backend directly bypasses the router’s direction filter and capacity gate.
  4. That roles are assigned narrowly; most people need operator, not admin.

Sources and references

Section titled “Sources and references”