Skip to main content

Firewall & Block Events

The Sielum Firewall lets you cut endpoints off from specific domains. Block rules are enforced at the OS network layer — connections to blocked domains are dropped before any data leaves the endpoint.

A block affects the whole device, not just the AI agent

A firewall rule is an OS-level rule keyed on the domain's IP addresses. It drops the connection for every process on the endpoint, including the user's own browser and tools that have nothing to do with AI. There is currently no way to scope a block to a single agent.

Use this to ban a prohibited service outright on a device. Do not expect it to isolate one agent while leaving the rest of the machine's network access intact.

Domain rules

Each firewall rule targets one domain (and all its subdomains) and has two switches:

FieldMeaning
EnabledWhether the rule is distributed to your agents at all
BlockWhether the rule is a block rule

Whether an enabled block rule actually drops traffic is not decided per rule — it depends on the enforcement mode of the endpoint, see below.

Audit vs. enforce is a per-device setting

Every endpoint runs in one of two enforcement modes. The default is audit, which is the safe setting for rollout.

ModeBehavior with an enabled block rule
Audit (default)Connection is not blocked; the match is logged as a Block Event
EnforceConnection is dropped at the OS firewall; the Block Event is logged

Set the mode in Devices: use the audit/enforce toggle on an endpoint row, or select several endpoints and apply the mode in bulk. The toggle is disabled for endpoints whose agent is not privileged — enforcement needs CAP_NET_ADMIN (Linux) or administrator rights.

This lets you write one set of rules and roll enforcement out device by device: keep most of the fleet in audit while you check the Block Events, then flip individual endpoints to enforce.

Creating a firewall rule

  1. Navigate to Firewall in the left sidebar
  2. Click New Rule
  3. Enter the target domain (e.g. api.openai.com)
  4. Optionally add a comment describing why the rule exists
  5. Click Save — the rule is created disabled
  6. Enable the rule when you are ready to distribute it

Enabled rules are picked up by each agent on its next check-in — by default every 5 minutes (see Agent Configuration). The agent resolves the domain to IP addresses and installs the corresponding outbound rules. Rules are removed again when the agent shuts down.

Example rules

DomainUse case
api.openai.comBlock direct ChatGPT API access from AI agents
api.anthropic.comBlock direct Claude API access (not via approved proxy)
generativelanguage.googleapis.comBlock Gemini API
copilot.github.comBlock GitHub Copilot completion endpoint

Block Events

Every time an AI agent's observed connection matches an enabled block rule, a Block Event is recorded — in audit mode as well as in enforce mode. Block Events are visible under Block Events and show:

  • Timestamp
  • Endpoint hostname
  • The process whose connection matched, and its PID
  • Destination domain and IP
  • Mode (audit / enforce)

The process and PID identify which agent's connection attempt matched the rule. In enforce mode the underlying block is still device-wide, so other processes reaching the same domain are dropped too, without appearing as separate events.

Block Events are preserved for audit purposes even after a rule is deleted or disabled.

How enforcement works

When the agent runs in privileged mode (requires CAP_NET_ADMIN on Linux), firewall rules are enforced via:

  • Linux: iptables (IPv4) and ip6tables (IPv6)
  • macOS: pf (handles both address families)
  • Windows: WFP (Windows Filtering Platform)

In user mode the agent cannot install firewall rules at all. Such endpoints stay in audit behavior no matter which mode is selected, and the enforce toggle is disabled for them in the dashboard. See Agent Installation → for details on operating modes.