Guides · ← All guides
AI Agent Egress Control: Allowlist Patterns That Work
By the Ingenious AI team · Published: November 6, 2026 · 8 min read
The short answer
The allowlist pattern that works is default-deny egress, bound per tool and per data class: the agent can send data only to destinations approved for that tool, carrying only the classes of data approved for that destination, with everything else refused or quarantined for human review. A single global domain list is the version that quietly fails — it ignores which tool is calling, what the payload contains, and the difference between your API and a lookalike subdomain. Start from the agent's real outbound history (a week of destinations, reviewed), write the list from evidence, and expand it deliberately: every addition is a small, logged policy decision, never an exception whispered to the model. Patterns as implemented in Rogue-Agent Defense's boundary engine, November 2026.
Pattern 1: default-deny, written from evidence
Begin in observe mode: log every outbound destination the agent actually uses for a week, review the list with a human, and promote exactly those destinations to the allowlist. Then flip to deny. Writing the list from imagination produces either a list so tight the agent cannot work (and the team disables the control) or so loose it approves the internet. Evidence first is what makes default-deny survivable — and the reviewed week doubles as your first rogue-agent baseline: any destination outside it, ever after, is a sign by construction.
Pattern 2: per-tool destination binding
One list per tool, not one list per agent. The email tool may reach your mail provider; the fetch tool may reach your documentation domains and your APIs; the webhook tool may reach exactly the endpoints it was configured for. A global list lets any tool use any destination — which means the fetch tool can POST to your mail provider's API with the agent's credentials, a combination nobody ever approved on purpose. Binding destinations to tools is the difference between an allowlist and a phone book.
Pattern 3: data-class rules on top of destinations
Where data may go depends on what the data is. Pair the destination list with taint classes: public content may flow anywhere on the list; internal documents only to your own domains; credentials and customer PII to no outbound destination at all, ever, through any tool. This is the layer that stops the "helpful summary" email — the destination (your mail provider) was allowed, the payload class was not. Destination-only lists cannot express that sentence.
Pattern 4: the quarantine lane
Binary allow/deny breaks on the gray cases: a legitimate new vendor domain, a one-off partner endpoint. Refusing them trains the team to bypass the control; allowing them trains the list to grow teeth-less. The working pattern is a third state — quarantine: the call is held, a human approves or rejects it once, and the decision is written back into policy. Gray traffic becomes policy input instead of either an outage or an exception. Our policy enforcement guide covers the approval mechanics.
Pattern 5: exact destinations, no wildcards that swallow the map
The two failure shapes to refuse at review: wildcard domains (*.example.com approves every subdomain, including the one an attacker registers or the forgotten test instance) and scheme-blind entries (approving a host for GET lookups and silently permitting POST uploads to it). Write entries as exact host + scheme + method where the tool supports it, and make every wildcard justify itself in the policy review like the exception it is.
Rollout order that survives contact with production
- Observe one week of real destinations; review with the agent's owner.
- Bind per tool; attach data-class rules to the sensitive tools first (email, fetch, file write).
- Quarantine instead of deny for two weeks; harvest the approvals into policy.
- Deny by default, quarantine lane kept; review the denied log weekly — it is now your earliest rogue-agent sensor.
- Plant canaries so a list failure announces itself the first time, not the tenth.
The context for all of it: egress control exists because the other layers — the model's judgment, the system prompt, output filters — can each be talked around. The exits cannot. The full threat picture is in how agents exfiltrate data.
Frequently asked questions
What is an egress allowlist for an AI agent?
The list of destinations an agent is permitted to send data to, enforced at the boundary with default-deny: anything not on the list is refused or quarantined. Written from the agent's real outbound history, bound per tool, and paired with data-class rules, it is the control that kills the most common exfiltration pattern outright.
Why does a single global domain list fail?
It ignores which tool is calling and what the payload contains. Any tool can use any approved destination, wildcards approve subdomains nobody reviewed, and a scheme-blind entry permits uploads where only lookups were intended. Per-tool binding and data-class rules are the fixes.
How do I build an allowlist without breaking the agent?
Observe first: log a week of real destinations, review them, promote exactly those, and run quarantine-instead-of-deny for two weeks so gray cases become human-approved policy instead of outages. Then flip to default-deny with the quarantine lane kept.
Should legitimate new destinations be added automatically?
No. Every addition is a small, logged policy decision by a human. Automatic expansion — especially anything the agent itself can request and receive — rebuilds the open egress the list exists to close.
What happens to calls that are not on the list?
They are refused, or held in a quarantine lane for one-time human approval where the workflow allows it. Either way they are logged: the denied log is your earliest warning that an agent has gone off-policy.
Test the Shield engine free → See egress and content controls score a prompt or agent output live.