MAQPNA

Learn

AI agent egress control and DLP.

An agent that can read your data and reach the internet can leak it. The fix is to control the way out, not to hope the model refuses.

Updated 1 November 2026.

FAQ

Common questions

What is egress control for AI agents?

A rule that an agent's environment can send traffic only through one governed exit, where each tool, model, web or agent-to-agent call is checked against policy. Everything else is denied by the network.

Can DLP stop prompt injection data theft?

Only partly. DLP catches recognisable patterns such as keys and card numbers. It misses paraphrased or encoded data. Pair it with taint-based rules that hold or deny outbound tools after the agent has read untrusted content.

Is a Kubernetes NetworkPolicy enough?

It is the foundation, if your CNI enforces it. It decides which hosts the pod can reach, but not what a call to an allowed tool contains. You still need a gateway that applies policy and DLP to each call.

Does MAQPNA send data to an external DLP service?

No. MAQPNA's DLP detectors run inside the gateway, and the audit ledger stores detector IDs and counts rather than the matched values.

Are model API calls egress?

Yes. Prompts can contain the same sensitive data as tool calls. In MAQPNA, model calls go through the gateway too, with DLP on prompts and completions, and can be pointed at local models.

Give agents one way out.

Run MAQPNA locally and watch a DLP profile redact a secret before it reaches a tool.

More install options
curl -fsSL https://maqpna.com/install.sh | sh