IRIS8 Talk to us

Secure

See and protect every AI request

Point your apps at one gateway endpoint to inventory AI use, mask sensitive data and block prompt injection inline.

Set-up

One endpoint, no rewrite

Change the base URL in your existing client and keep your code.

client setup
client = OpenAI(
  base_url = "https://your-gateway/v1",
  api_key  = env("IRIS8_KEY")
)

Works with OpenAI-compatible clients

Inventory

AI inventoryBuilt from gateway trafficSample data
  1. underwriting-copilotlendingActive
  2. support-copilotcontact centreActive
  3. coding assistantengineeringUnregistered

An inventory built from real traffic

Models, agents and tools register themselves the first time they call the gateway.

Masking

Sensitive data masked before it leaves

Personal data, secrets and source code are detected and masked inline, by rules you set per team.

Request logRules set per teamSample data
  1. Account numbersupport-copilotMasked
  2. API secret keydeveloper toolsMasked
  3. Source fileexternal toolAlert sent
  4. Prompt injectionemail agentBlocked

Findings

Gateway detections
  • Masked
  • Blocked
  • Alert sent
OCSF exportYour SIEM

Findings in the tools security already uses

Detections are exported in OCSF format to your SIEM.

Included

Everything included

Gateway

One drop-in endpoint

Inventory

A live map of models, agents and tools

Analytics

Requests, tokens and spend by team and model

Data masking

Personal data, secrets and code masked inline

Injection defence

Multi-turn attacks and poisoned tool results caught

Developer insight

AI use in engineering, seen from gateway traffic

Getting started

Three steps to start

  1. Point

    Point one workload at the gateway

  2. Tune

    Tune policies in alert-only mode

  3. Enforce

    Turn on enforcement, team by team

Questions

Questions

Talk to us
Q1Do we have to replace our existing gateway?

No. Secure runs as a drop-in endpoint.

Q2Does it slow requests down?

Checks run inline on the request path.

Q3Can we start in alert-only mode?

Yes. You can run alert-only first, then enforce per data class and per team.

Next step

See it on one workload