Platform
ScaiWave ScaiGrid ScaiCore ScaiBot ScaiDrive ScaiKey Models Tools & Services
Solutions
Organisations Developers Internet Service Providers Managed Service Providers AI-in-a-Box
Resources
Support Documentation Blog Downloads
Company
About Research Careers Investment Opportunities Contact
Log in

Network profiles

A bunker's network profile decides what traffic it can make. The five profiles range from "no network at all" to "full outbound plus L2 attaches", and each is gated by a separate permission so a tenant can grant exactly the network posture each user or agent should have.

The five profiles#

network_profile is set at create time on the bunker and can't be changed afterwards. Pick before you provision; if the posture needs to change later, take a snapshot, terminate, and create a new bunker.

isolated (default)#

No network. The bunker can talk to the controller's agent backchannel and nothing else — no DNS, no outbound TCP, no inbound anything. Best for running untrusted user-submitted code, evaluating model output, or any workload that genuinely doesn't need the internet. No permission required beyond scaibunker:create.

registry#

Allowlist of common package registries (PyPI, npm, crates.io, Maven Central, RubyGems, the Hugging Face Hub, and a handful of OS package mirrors). pip install, npm install, cargo build, apt update against the platform-managed apt mirror — they all work; nothing else does. Gated by scaibunker:network:registry. Use this for the common case of "I need to install deps but I don't want the bunker reaching arbitrary hosts."

allowlisted#

Tenant-configured domain list. You set network_allowlist on the bunker:

json
1
2
3
4
5
{
  "image": "python-3.12",
  "network_profile": "allowlisted",
  "network_allowlist": ["api.example.com", "*.internal.example.com"]
}

Allowlist entries are validated at the API edge:

  • Plain hostnames (github.com, pypi.org) and first-level wildcards (*.example.com) only.
  • No scheme prefixes, no path components, no double wildcards (**.foo.com), no mid-name wildcards (foo*.example.com).
  • Each label is 1-63 RFC 1035 characters.

An entry may name a TCP port: github.com:22. An entry without a port allows TCP 80 and 443; an address opens only on its entries' ports, and UDP is always dropped.

The worker enforces the allowlist via DNS interception plus per-flow connect filtering. An allowed name's addresses stay allowed for the bunker's lifetime, and TLS server names aren't checked, so other sites sharing a CDN address with an allowed name are reachable too. Gated by scaibunker:network:allowlisted.

unrestricted#

Full outbound. The bunker can reach anything its worker can, except private address ranges. Egress is audited as NDJSON batches surfaced through GET /bunkers/{id}/audit-batches. Gated by scaibunker:network:unrestricted — typically held by a small set of agents that need the full open internet (web-scraping agents, broad SDK testing).

transit#

L2-attached to one or more tenant-scoped bridges. Used to chain network appliances (firewall, IDS, NAT, router) at L2 in front of application bunkers — the same way you'd rack a hardware firewall in front of a server. Each interfaces[] entry names a bridge_name the bunker plugs into; the bunker gets one TAP per interface. Gated by scaibunker:network:transit.

json
1
2
3
4
5
6
7
8
9
{
  "image": "ubuntu-24.04",
  "network_profile": "transit",
  "interfaces": [
    {"if_name": "wan", "bridge_name": "tenant-wan", "spoof_guard": true},
    {"if_name": "lan", "bridge_name": "tenant-lan", "spoof_guard": true}
  ],
  "bandwidth_mbit": 100
}

Bridges are provisioned by tenant admins via POST /bridges. They are scoped to a tenant; a transit bunker can only attach to its own tenant's bridges (the worker also enforces this, defence-in-depth).

spoof_guard defaults to true and rejects packets whose source MAC isn't the interface's assigned MAC. Setting it to false is required for legitimate learning-bridge appliances and additionally requires scaibunker:l2_transparency; every such use is audit-logged.

Permission summary#

Each profile beyond isolated is gated by its own module-permission key. A caller without the key gets NETWORK_PROFILE_DENIED (403) at create time — the check happens before any worker round-trip.

Profile Permission required
isolated none (any caller with scaibunker:create)
registry scaibunker:network:registry
allowlisted scaibunker:network:allowlisted
unrestricted scaibunker:network:unrestricted
transit scaibunker:network:transit + bridges in this tenant

Adding spoof_guard: false to any transit interface additionally requires scaibunker:l2_transparency.

Bandwidth caps#

Every profile honours bandwidth_mbit, a per-bunker egress cap in megabits per second. The worker applies it via a TBF qdisc on the bunker's TAP. Acceptable range is 1-10000 Mbit/s; null means use the worker default (typically 100 Mbit/s). This is how tenants tier free vs paid plans — the same image and profile, but a tighter egress shaper for the free tier.

Choosing a profile#

A simple decision tree:

  • Does the bunker need outbound? No → isolated.
  • Does it only need package installs? → registry.
  • Does it need a small, known set of HTTPS endpoints (your API, your databases)? → allowlisted.
  • Does it need open-internet web access? → unrestricted, and turn on the audit batches.
  • Is it a network appliance plugged between tenants and other bunkers? → transit.

Default to the most restrictive profile that works. Most agents that need "the internet" actually only need registry plus one or two domains under allowlisted.

Where audit batches live#

Bunkers under unrestricted, allowlisted and registry have their egress audited: the worker writes per-flow records as NDJSON batches, kept 30 days. Each record's action is allow, or deny for a refused connection (at most 10 per second per bunker are recorded). Callers retrieve batches through GET /bunkers/{id}/audit-batches (list) and GET /bunkers/{id}/audit-batches/{name} (fetch).

Each batch is a {ts_us}-{worker_pid}.ndjson file containing one JSON object per flow: source IP and port, destination IP and port, hostname (if DNS-resolved), protocol, byte counts, duration. The natural lex order of the keys is chronological, so pagination is trivial — list with a since_us cutoff and follow next_continuation_token until done.

Changing posture later#

The profile is set at create time and can't be changed in place. If a bunker needs to move from isolated to registry, take a snapshot, terminate, and create a new bunker referencing the snapshot. This is by design: a profile change rewrites the worker's iptables and namespace setup for the bunker, and applying it to a running microVM has too many failure modes worth supporting.

Common mistakes#

  • Putting URLs in network_allowlist. Entries are hostnames, not URLs. https://api.example.com is rejected; api.example.com is accepted.
  • Double-wildcard. **.example.com is rejected — only first-level wildcards. Use one *.example.com entry per leaf domain, or list explicit hostnames.
  • Forgetting bandwidth_mbit on the free tier. A high-throughput bunker on the open internet can saturate a worker's uplink. Set a per-bunker cap on the bunkers you give to less-trusted tenants.
  • Granting scaibunker:network:unrestricted widely. Even with audit batches on, the unrestricted profile is the highest-risk profile by far. Grant it deliberately, per service account, and review the audit batches.

Transit-bunker scheduling#

The scheduler only considers workers that own every bridge a transit bunker references. If you create a transit bunker referencing tenant-wan and tenant-lan and no single worker has both, the call fails with NO_SUITABLE_WORKER. The fix is to create matching bridges on a common worker before launching the bunker. Bridge create is a two-phase operation — the controller reserves the row, calls the worker, and rolls the row back if the worker rejects, so the controller and worker ledgers never drift.

Transit bunkers can't be migrated between workers — the bridges they depend on are worker-scoped. Plan transit workloads with that in mind; for moveable workloads use one of the v0.1 profiles.

DNS#

Bunkers on registry, allowlisted and unrestricted reach the internet over both IPv4 and IPv6. Each gets a private address of each kind, and its outbound traffic leaves from its worker's public addresses. When a name has both, the bunker uses IPv4, so IPv6 carries traffic to IPv6-only destinations. The same blocks apply to both: private and other non-public ranges, other bunkers and the worker itself are unreachable, and traffic from an address that isn't the bunker's own is dropped. IPv6 flows are audited and count towards egress_bytes like IPv4. isolated bunkers have no network either way, and on transit the appliances configure their own addresses. Bunkers started before 17 November 2026 have IPv4 only until recreated. The bunker record's ipv6 field says whether a bunker has IPv6.

The registry and allowlisted profiles use a worker-side resolver that returns A and AAAA records only for hostnames the profile permits. NXDOMAIN comes back for everything else; there's no exfiltration channel through arbitrary DNS lookups. On unrestricted and transit, DNS resolution is unrestricted (and on transit, the bunker's own resolver may be a different appliance — e.g. a downstream firewall — depending on what the operator wires up).

The resolver lives on the worker and is reset between bunkers, so one bunker's leaks don't carry over.

Updated 2026-11-17 10:07:48 View source (.md) rev 43