---
title: "AI Firewall: one checkpoint for your AI traffic"
description: "Siper is an AI firewall on your own servers. It masks the personal data it detects, stops prompt attacks on the AI gateway, and logs each decision."
url: https://siper.io/ai-firewall
---

AI Firewall

# One checkpoint for your AI traffic.

Siper runs on your own servers, between your people, applications and agents and the AI services they use. It checks the traffic that passes through it on the way out of your network.

[Try the playground](https://siper.io/playground)

Where it sits

## Siper sits on the three paths AI traffic takes.

- Web proxy

  ### Employees

  Your web proxy decrypts AI traffic and hands it to Siper over ICAP. Siper masks typed prompts and uploaded files. Nothing is installed on employee laptops.
- AI gateway

  ### Applications

  Applications and SDKs point their base URL at Siper. The gateway is OpenAI-compatible and also serves the Anthropic Messages API.
- AI gateway and MCP

  ### Agents

  Coding agents such as Claude Code and Codex send their model traffic and MCP tools through Siper.

On each request

## Each request is masked, blocked or passed under your rules, and each decision is logged.

| Action | What happens |
| --- | --- |
| Mask | Each personal data value Siper detects becomes a placeholder before the request leaves your network. The AI service can still answer. |
| Stop prompt attacks | On the AI gateway, a detection model checks each request. It can block an attack, or record it and forward the masked request. |
| Block | By default, a request that Siper cannot mask is blocked. A blocked request can be sent again, and a request that left unmasked cannot be called back. |
| Log | Audit logs, web proxy decisions and AI gateway events go to your SIEM over syslog or Splunk HEC. Prompt text, file content and raw personal data stay out of them, and a proxy user is sent as a user name and domain. |

Files

## Uploaded files are masked in the same format.

On the web proxy path, Siper masks text, CSV, Word, Excel, PowerPoint and PDF files, and PNG and JPEG pictures. It reads scanned pages and pictures with text recognition and covers each value it finds.

A file that Siper cannot check, such as a password-protected PDF or an archive, is blocked. The AI gateway refuses files.

Limits

## Know where the firewall stops before you rely on it.

- Siper finds most personal data, not all of it. The [benchmark](https://siper.io/benchmark) gives the measured recall.
- Siper inspects the traffic that reaches it. Block HTTP/3, desktop apps that pin certificates and AI services outside your proxy rule on your network.
- Prompt attack detection runs on the AI gateway, not on the web proxy path.
- In pictures, Siper covers text it can read. It does not cover faces or signatures.
- If your proxy cannot reach Siper, your proxy's own setting decides. Set it to fail closed.
- Monitor mode, report mode, API keys with masking turned off and emergency mode let data pass on purpose. Each one is visible in the panel and in the audit log.
- On developer APIs, formats the web proxy cannot read pass unread and are reported.

Next step

## See what Siper finds in your own text.

[Book a demo](https://siper.io/contact)[Try the playground](https://siper.io/playground)
