Agent DNS

Your AI agents can reach anything.

Give them only what they need.

Securd Agent DNS gives each agent its own resolver policy. Approved destinations resolve. Anything new is held for review. It runs in the network, so an agent that goes off script cannot switch it off.

Deploys as a resolver setting. No SDK, sidecar or proxy.

Agent Connection Monitor
Query
Compartment
Scope
First-Seen
Feeds
Policy
Answer
Event
api.openai.com Approved
mcp.unknown-tool.io Held
oai.azure.com Approved
api.newai-svc.xyz Held
api.serper.dev Approved
exfil-data.click Blocked
api.anthropic.com Approved
rogue-api.ru Threat
ApprovedBlockedHeldThreat

Sample agent connections

The Problem

Open DNS is an exit path
for a compromised agent

Every tool call, model call and data fetch begins with a DNS lookup. An agent with an unrestricted resolver can reach any destination a prompt injection, poisoned tool description or compromised dependency names. Blocklists do not cover destinations that did not exist when the list was built. Egress firewalls cannot attribute a connection to a specific agent.

Architecture and limits

Scope per policy

Each agent is assigned a policy that resolves only its approved destinations. An invoice agent resolves the ERP and the payment API and nothing else.

First-seen names are held

The Greywall holds names the policy has not previously resolved. Held names are approved, blocked, or released when the hold expires.

Enforced outside the agent

Policy is applied by the resolver, not by code in the agent runtime. A compromised agent receives the same answers as an intact one.

How It Works

Deployment in three steps

A resolver setting, a period of observation, then policy enforcement. No SDK in the agent and no proxy in the data path.

STEP 01

Assign the resolver

A VPC DHCP option, a CoreDNS forward, a dnsConfig on the pod, or a DoH URL in the runtime. No software is installed.

STEP 02

Review the inventory

The policy records every hostname the agent resolves: model APIs, MCP servers, tool endpoints and data sources. Learning mode records first-seen names without holding them.

STEP 03

Enforce the policy

Approve required destinations, block the rest, and enable the Greywall hold for new names. Every decision is logged and forwarded to your SIEM.

Scope of the control

Securd operates at DNS resolution only

Securd evaluates one question for each lookup: whether the requesting policy is permitted to resolve the requested name. This is why deployment is a resolver setting and why the control is not affected by changes inside the agent runtime.

No TLS inspection

No certificates to distribute. Applications that pin certificates are unaffected.

No host software

The same resolver setting applies to containers, functions, workstations and CI runners.

No proxy in the data path

Traffic is not routed through Securd. The only added latency is DNS resolution.

The limits of a DNS-layer control and the accompanying egress rules are documented on the Why Agent DNS page.

One policy per agent

An invoice agent requires the ERP API and the payment processor. A research agent requires a model provider and a search tool. Neither requires access to the rest of the internet, and neither should be able to reach it on the basis of an instruction in retrieved content.

Deployed as a resolver setting with no software in the agent.

First-seen names are held; learning mode records them during initial deployment.

Every decision logged and forwarded to your SIEM.

Common Questions

What is Securd Agent DNS?

A recursive DNS resolver with a policy per AI agent, environment or team. Each policy resolves only the destinations it permits. First-seen names are held at the Greywall. Enforcement is performed by the resolver, outside the agent runtime, and deployment requires no software in the agent.

How does the Greywall work?

It holds any name a policy has not previously resolved, for a configured hold time. One policy per agent gives each agent an independent baseline. Learning mode records first-seen names without holding them during initial deployment. Held names are approved, blocked, or released according to policy; a held name is not allowed by default as it ages.

How is this different from a newly-seen-domains category?

Those categories are global and derived from the vendor's telemetry. The Greywall baseline is maintained per policy inside your compartment, the hold duration is set by the policy, and enforcement is performed outside the agent. A feature-by-feature comparison is on the Why Agent DNS page.

What does an agent see when a destination is held?

A DNS answer containing the block page address configured for the policy, rather than a timeout. Application code can distinguish a held destination from an unavailable service by comparing the answer against that address. Reference implementations are in the developer library.

How is Securd priced?

Annual plans measured in compartments, policies, deduplicated events, storage and forwarding. Agent identities are unlimited. MSSP plans are priced per tenant. Plan details are on the pricing page.

Where do the logs go?

Events forward over syslog, export as CEF, JSON or CSV, and are available through the API and as six signed webhook events. A push API accepts verdicts from a SOAR. Configuration changes are written to a change log with the user, timestamp and content.

Evaluate Agent DNS with your own agent traffic

Deploy on a single policy in learning mode and review the results with your security team.