Skip to content
Rexaris AI

LLM firewall: inspect every AI request before it leaves your perimeter

A network firewall decides which packets may pass. An LLM firewall decides which prompts, data and responses may pass between your people and AI models. Rexaris AI runs that decision inside your infrastructure.

Employees, applications and agents now send text, files and code to language models all day. Most of that traffic is harmless. Some of it carries customer records, credentials or confidential documents, and nothing in a conventional proxy understands the difference.

Rexaris AI sits between your users and the model providers as an OpenAI- and Anthropic-compatible gateway. Every managed request is classified, matched against your policies and allowed, masked, blocked or routed, with the decision written to an audit log before the response is returned.

Request a Private Pilot

What a web proxy cannot do

  • Understand content

    A proxy sees a POST to api.openai.com. An LLM firewall sees a card number inside a prompt, an API key inside a file attachment, or a confidential label on a document.

  • Act per data class

    The same request can be masked for one destination and blocked for another. Policies are written in terms of data classes, users, groups and providers, not IP ranges.

  • Check the response path

    Model output is inspected too: secrets that come back in a completion can be masked before a developer or an application sees them.

  • Stay deterministic

    Rules fire the same way every time. No model is asked whether a request looks acceptable, so an auditor can reproduce every decision.

How Rexaris AI works as an LLM firewall

  • Sensitive data detection

    PII, payment data, credentials, secrets, source code and your own patterns are detected on the request and the response.

    Learn more

  • Policy engine

    Conditions on user, group, data class, application and provider resolve to Allow, Mask, Block or Route.

    Learn more

  • Prompt and content protection

    Prompt-injection and jailbreak patterns, prohibited topics and custom keyword lists, with a clear statement that detection is one layer, not a guarantee.

    Learn more

  • Audit trail

    Who sent what, which rule fired, which model answered, how many tokens it cost.

    Learn more

Frequently asked questions

Does an LLM firewall replace my DLP or SASE product?
No. Those tools see traffic to any destination from managed endpoints. Rexaris AI understands the content of AI requests that pass through its gateway and enforces policy there. The two work together; Rexaris can also ingest shadow AI events from them.
Will it slow down responses?
Classification runs inside your network before the request is forwarded. Typical overhead is tens of milliseconds, far below model latency. Masking and restoring placeholders adds no round trips.
Can it block everything that is sensitive?
It can block what its detectors and your rules identify. Detection is pattern- and classifier-based and should be combined with routing confidential work to internal models rather than relied on alone.
Where does it run?
Entirely inside your infrastructure: Docker or Kubernetes, your database, your audit logs. Customer prompts do not pass through our cloud.

Related solutions

From the blog

See your own traffic through an LLM firewall

Start with one team in audit-only mode. The data classes in real prompts will tell you which rules to write first.

Request a Private Pilot