greywall/docs/security-model.md

Security Model

Fence is intended as defense-in-depth for running semi-trusted commands with reduced side effects (package installs, build scripts, CI jobs, unfamiliar repos).

It is not designed to be a strong isolation boundary against actively malicious code that is attempting to escape.

Threat model (what Fence helps with)

Fence is useful when you want to reduce risk from:

• Supply-chain scripts that unexpectedly call out to the network
• Tools that write broadly across your filesystem
• Accidental leakage of secrets via "phone home" behavior
• Unfamiliar repos that run surprising commands during install/build/test

What Fence enforces

Network

• Default deny: outbound network is blocked unless explicitly allowed.
• Allowlisting by domain: you can specify allowedDomains (with wildcard support like *.example.com).
• Localhost controls: inbound binding and localhost outbound are separately controlled.

Important: domain filtering does not inspect content. If you allow a domain, code can exfiltrate via that domain.

How allowlisting works

Fence combines OS-level enforcement with proxy-based allowlisting:

• The OS sandbox / network namespace is expected to block direct outbound connections.
• Domain allowlisting happens via local HTTP/SOCKS proxies and proxy environment variables (HTTP_PROXY, HTTPS_PROXY, ALL_PROXY).

If a program does not use proxy env vars (or uses a custom protocol/stack), it may not benefit from domain allowlisting. In that case it typically fails with connection errors rather than being "selectively allowed."

Localhost is separate from "external domains":

• allowLocalOutbound=false can intentionally block connections to local services like Redis on 127.0.0.1:6379 (see the dev-server example).

Filesystem

• Writes are denied by default; you must opt in with allowWrite.
• denyWrite can block specific files/patterns even if the parent directory is writable.
• denyRead can block reads from sensitive paths.
• Fence includes an internal list of always-protected targets (e.g. shell configs, git hooks) to reduce common persistence vectors.

Visibility / auditing

• -m/--monitor helps you discover what a command tries to access (blocked only).
• -d/--debug shows more detail to understand why something was blocked.

Limitations (what Fence does NOT try to solve)

• Hostile code containment: assume determined attackers may escape via kernel/OS vulnerabilities.
• Resource limits: CPU, memory, disk, fork bombs, etc. are out of scope.
• Content-based controls: Fence does not block data exfiltration to allowed destinations.
• Proxy limitations / protocol edge cases: some programs may not respect proxy environment variables, so they won't get domain allowlisting unless you configure them to use a proxy (e.g. Node.js http/https without a proxy-aware client).

Practical examples of proxy limitations

The proxy approach works well for many tools (curl, wget, git, npm, pip), but not by default for some stacks:

• Node.js native http/https (use a proxy-aware client, e.g. undici + ProxyAgent)
• Raw socket connections (custom TCP/UDP protocols)

Fence's OS-level sandbox is still expected to block direct outbound connections; bypassing the proxy should fail rather than silently succeeding.

Domain-based filtering only

Fence does not inspect request content. If you allow a domain, a sandboxed process can still exfiltrate data to that domain.

Not a hostile-code containment boundary

Fence is defense-in-depth for running semi-trusted code, not a strong isolation boundary against malware designed to escape sandboxes.

For implementation details (how proxies/sandboxes/bridges work), see ARCHITECTURE.md.