FICCDocumentation
nightly prerelease / Version 0.3.0.dev1Project website
Reference library
Reference
On this page
  1. Policy packages

Security#

This policy covers the 0.2.6 stable release series, including revision releases. Report a security issue privately to contact@federatedindustrial.com. Do not include credentials, private keys or confidential logs in a public issue.

Local mode uses loopback access. Do not expose its listener through a public proxy, tunnel or network bind. Stable releases also provide explicit remote mode: verified HTTPS, a protected Unix gateway connection, approved external identities and current project permissions. Adding a proxy to a local-mode listener does not enable secure remote access.

The local account, its SSH configuration and its SSH agent are trusted. A process with the same account can control FICC and use that account's keys. The service does not isolate applications that share the local account.

Application identities separate project workspaces and operation receipts within that trusted account. The owner controls membership and can issue credentials through the private socket. The owner assigns project machines, folders and capabilities. Job and file receipts remain project-bound; terminals also require their creating identity. Job execution and shells have full remote-account authority. Provider and installation administration remain owner-only. Remote identity approval is separate from membership and cannot assign the local recovery owner. See identities and projects for ownership and recovery limits.

Remote machine data is untrusted. The service validates data before storage. The browser treats names, errors and resource labels as text. SSH connections require a pinned, trusted host key. A changed host key prevents collection.

Only locally approved SSH profiles can be enrolled. Review their proxy commands and executable configuration rules before approval. FICC does not accept SSH configuration text from the browser or store SSH private keys.

Browser sessions use short-lived login credentials and HttpOnly cookies. Origin, Host and CSRF checks apply to browser requests. API tokens have explicit scope, node/root access and expiry. Revoke unused credentials in Access or the CLI.

Managed execution has the full authority of the remote SSH account. Limits and GPU visibility are resource controls, not a sandbox for hostile programs. Grant jobs:execute only to callers permitted to use that account. Log output can carry secrets and needs the separate jobs:logs scope. Command arguments and explicit environment values are stored in durable job requests; avoid embedding secrets.

The CLI stores private submission credentials only in bounded recovery receipts under the state directory. They expire after one hour and remain revocable. Copy those receipts only with the same care as API credentials. Do not expose one-use browser bootstrap URLs through desktop configuration or service logs.

The service stores inventory and audit data in its private state directory. Do not put that directory in a repository or a shared filesystem. A private Git repository is not a suitable secret store. Local audit data is not tamper-proof against the account owner or system administrator.

Dependency checks and a source scan run before integration. These checks do not replace code review or establish that software has no security defects.

File roots are object capabilities within the trusted local or remote account. They refuse symlink traversal and bind registered directory identity. They do not isolate another program with the same account. An opened directory remains usable after external movement. Hard links can give an object other names.

Keep root namespaces operator-managed. Detected changes are refused; continuous pathname confinement against a concurrent same-account writer is not promised.

Interactive terminal execution has full remote-account authority. Ticket, Origin and current-grant checks protect attachment; revocation detaches the client. Persistent tmux sessions require a separate confirmed stop. Terminal bytes are untrusted, confined to the emulator and not recorded by FICC.

The browser content policy permits inline styles for xterm's generated font, RGB color and contrast rules. Scripts remain restricted to local assets. Application code constructs labels with text nodes and does not render remote HTML. Terminal color values are parsed by the pinned emulator; remote terminal data does not become arbitrary CSS or HTML.

Coding-agent launch and bus send have separate node-scoped permissions. Selecting a runtime profile does not change provider login, model settings, sandbox settings or approval policy. Direct delivery can start work under the recipient agent's existing account. Bus messages are labeled participant testimony and carry no FICC approval authority. Revocation prevents later admission but cannot undo work that the runtime already accepted.

The agent observation interface releases permitted machine inventory and resource samples under separate observations:read and observations:resources permissions. Current project machine assignments, credential limits and policy grants still apply. Grant these permissions only when the receiving agent and its inference provider may receive the documented observation fields. Revocation prevents later reads; it cannot recall information already released.

Observation tools do not execute commands, read files or retrieve logs. Remote credentials stay with the controller's existing transport. The observation client uses a separate API credential restricted to observation scopes. The endpoints refuse broader credentials and browser sessions. This interface does not constrain an agent's independently available tools, credentials or network access.

Nodes expose only the fixed SSH helper exchange and private local spool tools. Controller credentials stay on the controller. An outbox cannot choose a different sender, run or grant. Runtime submissions with uncertain outcomes are not replayed automatically.

Ordinary same-account processes remain inside the trusted account boundary. Spools reject links, unsafe ownership, unknown members and capacity overflow. The native TUI owns runtime approval decisions.

Native macOS controllers use kernel peer credentials for the private control socket and per-user launchd startup. Mac nodes support resource observations, terminals and coding agents. Managed resource-limited jobs and safe file roots still require Linux kernel facilities and report unavailable on Mac nodes. Agent shells and tools retain the runtime account's authority on either system; the macOS port does not add an agent sandbox. See macOS support for supported capabilities and qualification limits.

Runtime modules have a separate containment boundary. Executable packages require private namespaces, syscall filters and enforced cgroup limits. They receive no SSH keys, controller database, host home, display socket or provider socket. The host broker intersects current caller access, package-digest grants and panel targets. Missing containment prevents execution; there is no unrestricted fallback.

Package inspection validates bounded contents and hashes. It does not authenticate publishers. All current packages show an unverified publisher state and need explicit source acceptance. New digests need new grants. Disabling a package revokes its calls and streams while retaining recovery records and saved data.

Native displays use a separate verified runtime in their own bounded process. Opaque references bind a display to the current actor, panel and VM identity. Input is released on arrival and on focus or authority loss. Clipboard, file transfer, console audio and microphone input are disabled. See display controls.

Workspace data has local-account file protection, not encryption at rest. Audio uses an explicit browser-local file selection and revocable playback leases. No module may place provider passwords in workspace state. Treat saved notes, operation history and backups as private data.

Policy packages#

Signed policy archives contain Rego and JSON. Exact manifest signatures and payload hashes bind them to an explicitly trusted Ed25519 publisher. Ordinary workspace modules cannot establish that trust or install a security provider. Providers are separately installed trusted Python components with controller authority. The OPA provider requires a network-isolated process and verified resource controls for both compilation and evaluation.

Policy input contains current host-established authority, with caller labels kept separate. An allow can only narrow host grants. Invalid, missing, stale or unavailable required decisions deny access. Unrestricted local owner policy administration remains available for explicit recovery and generates audit events. Restore disables policy execution and publisher trust. See policy operations for the remaining managed-process limits.

FICC / DocumentationBack to top

Search documentation

Press Escape to close

Loading search...

    Reading appearance

    Choose colours, surfaces and text independently. These settings are saved for this documentation site in this browser.

    Flat surfaces remove decorative materials. Typography is a separate choice. Increased contrast uses the Contrast palette. System forced colours always take precedence.