Enforce cloud computer identity, secret, egress, and residue isolation #10

Open AtlantisPleb opened this 6h ago

Project

Cloud computer platform

Source: Cloud computer scale architecture audit

Outcome

Define and enforce one cross-provider security contract for cloud computers. A runtime must receive only the capabilities for one owner, workspace, lease, command, and generation, and it must produce verifiable zero-residue evidence before capacity returns to service.

Threats

  • Guest-to-host or sandbox-to-node escape.
  • Cross-tenant filesystem, process, network, checkpoint, event, or credential access.
  • Google metadata or service-account credential theft.
  • Confused-deputy provider routing and stale-generation mutation.
  • Unauthorized package, source control, model, or arbitrary network egress.
  • Secret capture in checkpoints, logs, command output, crash data, or warm runtimes.
  • Cleanup failure followed by unsafe capacity reuse.
  • Control-plane credentials or topology entering agent-visible payloads.

Scope

  • Mint short-lived, capability-scoped, generation-bound runtime credentials and remove bootstrap secrets after reverse-dial connection.
  • Give guests no ambient Google service account, token scope, project SSH key, or provider administration handle.
  • Block the metadata service and default-deny ingress and egress.
  • Route admitted source control, package, model, artifact, and control traffic through an observable policy broker.
  • Encrypt checkpoints with workspace-scoped keys and exclude runtime credentials and transient state.
  • Separate runtime hosts and clusters from Phoenix, PostgreSQL, control credentials, and customer-connected-computer controllers.
  • Enforce owner, tenant, workspace, path, symlink, process, artifact, output, and byte boundaries below every adapter.
  • Produce cleanup evidence for processes, mounts, overlays, devices, network rules, cgroups, credentials, scratch data, and checkpoints.
  • Quarantine a lease, sandbox, or host when cleanup cannot prove zero residue.
  • Redact raw provider payloads, private paths, addresses, tokens, and topology from logs, events, issues, and receipts.

Deliverables

  • Cross-provider threat model and security profile.
  • Runtime credential broker and rotation or revocation path.
  • Network and metadata policy broker.
  • Checkpoint key and redaction policy.
  • Cleanup proof and quarantine state machine.
  • Cross-tenant, escape, exhaustion, revocation, egress, secret-leak, and residue tests.

Acceptance criteria

  • A compromised runtime cannot obtain host or Google credentials or affect another tenant.
  • Revocation blocks new effects immediately and limits an in-flight command to its safest bounded checkpoint.
  • Unauthorized egress fails with a typed policy result and an operator-visible receipt.
  • No secret appears in a checkpoint, retained event, artifact metadata, public error, or warm runtime.
  • Failed cleanup prevents capacity reuse and opens an incident with bounded public-safe evidence.
  • Every accepted provider passes the same security corpus before production routing can select it.

Dependencies

Builds on #3 and supplies a required gate for every provider issue.

  1. AtlantisPleb opened this issue 6h ago
Sign in with GitHub to comment on this issue.