Your AI gateway is not a force field
by Jean-Philippe Evrard (employee post)A LinkedIn sponsored post takes a $2 DNS ad blocker and turns it into a lesson about enterprise AI governance. "That is how you control AI in a company too."
The proposed approach combines a central enforcement point, decision logging, and automatic blocklist updates from a remote source.
My objection is about the post's tone: It presents this solution as a panacea. A gateway can enforce policy on the requests it handles. If the advice encourages people to assume it also covers actions outside that path, it risks creating a harmful and false sense of security.
What the DNS blocker actually controls
A DNS blocklist filters queries sent to the resolver using it. If users can choose another resolver, they can avoid that filtering.
Preventing that requires controls elsewhere, such as network restrictions or device management. The dongle can be part of that architecture, but plugging it in does not establish the enforcement boundary.
Consider a newly registered domain hosting a DNS-over-HTTPS endpoint. If the domain is absent from the blocklist and the surrounding controls allow the connection, a client can use it for DNS resolution. The blocklist does not need to malfunction for this to work.
DNS can also carry communication itself, including traffic used for command and control or data exfiltration. That raises a separate question from which resolver a client uses. See MITRE ATT&CK T1071.004 for an explanation of DNS as an application-layer protocol used for command and control, and T1041 for exfiltration over an established command-and-control channel. These are distinct behaviours: resolving a destination, communicating with an attacker, and moving data.
Your dns-based ad-blocking device will not protected you from exfiltration of sensitive data.
Owning the chokepoint in AI?
An employee or agent calls an approved model through the approved gateway. The gateway authenticates the caller, checks the model, records the request, and allows it.
The agent then uploads sensitive material to an external endpoint using a tool it can access. Was the model call governed? Perhaps. Was the data transfer governed? That depends on the permissions and controls at the tool, the data source, and the destination. The model gateway cannot automatically confer control over those other systems. You have to build and enforce that control.
I find security frameworks useful here because they make the scope of enforcement explicit.
NIST SP 800-53 helps me turn that scope into concrete questions:
- AC-3, Access Enforcement: Which approved authorisations does the system actually enforce?
- AC-4, Information Flow Enforcement: Which flows of information are permitted, and where are those restrictions enforced?
- SC-7, Boundary Protection: Which communications cross the boundary, and which interfaces control them?
- AU-12, Audit Record Generation: Which events produce records?
Notice that generating an audit record is a different control from enforcing an information-flow restriction. A perfectly logged leak is still a leak.
Who can change the policy?
I agree that policy changes need authentication and an audit trail. A control whose configuration anyone can change is hard to trust.
But protecting the settings is only part of the argument. I still need to know what the policy permits, which paths it covers, and whether an approved destination can be used in a harmful way.
Updating the list does not define the policy
Automatic updates address the maintenance of the blocklist. They do not settle what your company should allow.
A domain's classification does not tell me whether a particular transfer of customer records or source code is authorised. That decision needs context the list does not contain.
Where the gateway helps
An AI gateway can be a useful enforcement point. Centralising model access can make credentials, request restrictions, and audit records easier to manage. That is a substantial benefit.
My objection is to treating that benefit as evidence of broader coverage. If the claim includes preventing sensitive data from leaving, I need to see how the controls govern the relevant tools, data sources, and destinations.
I would accept that claim if the architecture and testing demonstrated the coverage. The presence of a gateway alone does not demonstrate it.