LATESTBeyond AI Safety: The Business Rules Your AI Agents Must Not Break
An AI agent can be safe, helpful and completely wrong for your business. Enterprises also need to enforce their own business rules.
Your AI rules,on every prompt and every answer.
Noozle Gate integrates with your AI gateway to enforce security and business policies across employee AI conversations and agent tool calls. Keep sensitive data within approved boundaries, with readable policies your team can audit and interactions you can trace.
A policy check on the path of every request.
Gate runs alongside your AI gateway, enforcing your policies across models from OpenAI, Anthropic, Google Gemini, DeepSeek and other providers. It checks requests before they reach the model and responses before they reach the user, so switching models doesn’t mean changing your rules.
Users, apps, agents
Chat tools, internal apps and AI agents
Your AI gateway
One entry point for every model call
LiteLLM · full integrationAny LLM provider
Hosted, EU-private or self-hosted models
checks the requestchecks the answer
Noozle Gate
Your policies, evaluated. EU cloud or on-premise.
Decision log
IDs, policy revisions and outcomes, for your audit trail.
Streamed answers are checked as they flow.
Works with your AI gateway
- LiteLLMFull integration
- agentgatewayBeta
- Kong AI GatewayBeta
- BifrostBeta
No AI gateway yet? Start with LiteLLM.
Start LiteLLM with your provider credentials.
Connect Gate using the configuration supplied for your pilot.
Connect an app and verify an allowed request and a blocked request.
Using a different gateway? Tell us which one, and we will gladly integrate with it.Tell us your gateway
Add every policy you need. Your AI stays fast.
As your business grows, your policies grow with it. Gate’s highly optimized policy evaluation keeps AI interactions fast while applying your rules accurately. Your team can expand security controls and business rules with confidence in both speed and correctness.
1 · The request
team sales · model public-llm
“Draft a reply that beats the 18% discount Acme offered them.”
Gate checks every policy that could apply to this request, and only those.
2 · Your policies
7 could apply, checked in full · 1 match
Can’t apply · ruled out
and thousands more
Illustrative example.
Autonomous agents follow your rules too.
An agent works through the same model calls as a person, only faster and without anyone watching. Gate checks every step it takes, so the rules you write for people also apply to your agents, along with rules written just for them.
- 1 · Instructions
What the agent is told
Check the instructions and the context an agent starts from, for example which data a support agent may bring into a conversation.
Checked by Gate - 2 · Tools offered
What the agent may use
Decide which tools each agent or team may be given. A support agent is never offered a payment tool.
Checked by Gate - 3 · Tool calls
What the agent is about to do
Check the actions the model asks the agent to take, including names and arguments, before they run. Stop a refund above a limit or a message to an outside address.
Checked before it runs - 4 · Results
What comes back
Apply rules to what tools return into the agent's context, such as documents and web pages, before the next model call.
Checked by Gate
The result goes back to the model and the loop starts again. Gate checks every turn.
Gate decides on every agent step that passes through your AI gateway, while your tools keep running where they run today. Every stopped step is logged with the agent's identity, the policy and the revision.
One policy language, from exact patterns to meaning.
Your team writes the rules, for security and for your own business, in one short and readable policy language. Every change creates a new revision. A privacy officer can read a policy, an auditor can review it, and every decision points back to the exact policy and revision that made it.
Deterministic
Patterns, fuzzy and proximity matching, and comparisons on any field of the request.matches · fuzzy · near
Runs first
Linguistic
Language detection, lemmas and proper nouns, so a rule follows every form of a word.detect_language · lemma · propn
Runs next
Semantic
Multilingual similarity to a topic described in plain words, and key-passage extraction.about · about_score · textrank
Runs last, only when needed
Every rule can also use context: the key, team or user making the call, the model, the tools offered or called, and whether it is the request or the response. The bars show relative cost: semantic checks run only when cheaper conditions already hold.
Guardrails give you a score. Gate shows you the rule.
Many LLM guardrails give you fixed categories to switch on or off, and when something is blocked you see a label and a score. With Gate every policy is yours: written in your words, adjustable down to the last condition, and explained on every decision.
A typical guardrail
- Personal data
- Toxicity
- Prompt attack
- Custom topics
Blocked · category personal_data · score 0.91
Fixed categories, a threshold, and a verdict you cannot open.
A Noozle Gate policy
// Health data stays on EU-hosted models !(doc.model startsWith "eu-private/") your models && doc.request_data.team != "team-claims" your exception && about(text, "patient diagnosis", 0.85) your threshold
Blocked · health-topics-eu-only rev 7 · every condition traced
Every condition is readable, editable and versioned by your team.
| Can your team… | Typical guardrails | Noozle Gate |
|---|---|---|
| …read and change what a policy checks? | Usually a fixed category to switch on or off | Yes, every condition is yours |
| …see why a request was blocked? | Usually a category label and a score | Yes, the full reasoning with the matching text |
| …make exceptions ("everything except…")? | Usually left to a model to interpret | Yes, written into the rule |
| …match competitor or product names? | Usually described in words for a model to interpret | Yes: names, patterns, fuzzy matching |
| …depend on team, user, model and tools? | Usually set per configuration | Yes, in the same rule |
| …get the same decision for the same input? | Varies when a model is the judge | Yes, traced to a policy revision |
A category comparison. Individual products differ, and fixed categories are often quicker to switch on.
Policies as code, your way.
Manage Gate policies with the Noozle Terraform provider or the command line, in the same repositories and pipelines as the rest of your infrastructure.
- Reviewed like code. Every policy change is a pull request, approved by security, privacy or legal before it goes live.
- One process for every team. Plan, apply and roll back policies with the pipeline you already trust.
- Scriptable from the command line. Validate, dry-run and publish policies with
noozle-cli, from a laptop or any CI job. - Test, then promote. Run the same files in test and production, so what you approved is what enforces.
- An audit trail for free. Your git history shows who changed which rule, when and why.
- Built for scale. Hundreds of teams and policies from one place, with drift shown on every plan.
resource "noozle_query" "health_eu_only" { name = "Health data stays on EU-hosted models" description = "Owned by compliance. Reviewed quarterly." full_expression = <<-EOT !(doc.model startsWith "eu-private/") && about(text, "patient diagnosis", 0.85) EOT tags = ["gdpr-art-9", "compliance"] is_active = true }
$ terraform plan~ noozle_query.health_eu_only threshold 0.80 → 0.85Plan: 0 to add, 1 to change, 0 to destroy.
$ noozle-cli queries validate --expr-file health.expr $ noozle-cli queries create \ --name "Health data stays on EU-hosted models" \ --expr-file health.expr --direction request
Every decision can be traced. Failures are never silent.
MATCH q:377 rev 4 request
Keep pricing and discount terms out of external AI
Query reasoning
AND
metadata request_data.team == "team-sales" matched
OR
semantic about(confidential discounts and commercial terms) score 0.8712 / 0.8500
regex (floor price|discount) skipped after ancestor short-circuit
keyword "Acme" matched
Decision log: block, rule_match, q:377 rev 4- Explain any decisiona trace shows which conditions ran, each score against its threshold, what was skipped and the text that matched.
- A known violation always blockseven when another rule errors or times out.
- Incomplete evaluations are labelledand follow the fail-open or fail-closed setting you choose.
- No accidental bypassa disabled Gate or suspended tenant is rejected, not waved through.
- Safe policy changesnew rules are validated without calling a model; if a revision fails, the last valid one keeps enforcing.
- Correlated with your gatewayrecords carry your gateway’s call and trace IDs for incident review.
Always. Every decision is logged with its policy and revision.
On demand. Open a trace session for a policy, for a limited time, and authorized people see the full reasoning, including the matching snippets. Traces are streamed live, straight to the people who asked for them.
Start in our EU cloud. Move on-premise when you are ready.
The same product and the same policy language in both. Every pilot starts in the cloud, so you can try Gate before installing it yourself.
EU cloud
Made and hosted in EuropeEvery pilot starts here.
- Operated by Noozle on EU-owned infrastructure
- Prompts and answers processed in memory and discarded
- Decision records kept for your audit trail
- Lower price, suited to smaller companies and pilots
On-premise
When you are ready to run it yourself.
- In your data centre or private cloud
- Everything stays inside your perimeter
- You set log retention
- Suited to large regulated organisations
Never stored
Prompt and answer text, tool definitions and tool-call arguments. They are processed in memory and discarded.
Kept for your audit trail
Evaluation ID, policy and revision, outcome and completeness, model name, gateway call and trace IDs, timings.
More people using AI. More prompts. More policy violations.
Company data is flowing into AI. Sensitive data is at risk in both prompts and answers.
45%Increase of 200%vs. 2025 report
of employees regularly use AI on corporate devices.
2025 report: 15% Verizon DBIR, 202669per weekIncrease of 57%vs. 2025
data policy violations involving sensitive data sent to AI, in the average organisation.
2025: 44 per week Netskope AI Report, 202631per weekIncrease of 158%vs. 2025
data policy violations involving sensitive data returned by AI.
2025: 12 per week Netskope AI Report, 2026Relative increases within each source’s dataset, rounded to the nearest whole percent. Verizon compares its 2026 and 2025 reports; Netskope compares 2026 observations with 2025. Regular AI use means access at least once every 15 days.
Netskope also reports a sixfold increase in monthly prompts sent to generative AI apps.
Written policies set the rules; enforcement makes them hold. Nearly half of the privacy and security professionals surveyed by Cisco admit to entering personal or non-public data into GenAI tools (Cisco, 2025). Safety research points the same way. Aligned models can still be jailbroken (JailbreakBench), and detectors both miss attacks and block legitimate work (PIGuard, CAPTURE, GuardBench). That is the case for an independent layer, with rules you can read, that checks what goes in and what comes out.
Technical controls for GDPR and the EU AI Act.
Gate turns your compliance decisions into enforced, logged controls, and gives your DPO evidence for each one.
| Control | How Gate implements it | Supports |
|---|---|---|
| Personal data and identifiers | Pattern, fuzzy and semantic rules on prompts | GDPR Art. 5, 25, 32 |
| Special-category data | Multilingual topic rules by team and model | GDPR Art. 9 |
| Secrets and source code | Pattern rules on requests and responses | GDPR Art. 32 |
| Tool and agent use | Rules on tools offered and tool calls | OWASP LLM: excessive agency |
| Traceability | Decisions tied to policy revisions | AI Act Art. 12, 26 |
| Prompt injection | Rules on requests, answers and tool calls | AI Act Art. 15 · EDPS guidance |
What Gate does today, and what comes next.
-
Today
Enforcement
Allow or block requests and responses. LiteLLM fully integrated; agentgateway, Kong AI Gateway and Bifrost in beta.
Rules
Deterministic, linguistic and multilingual semantic rules on text, tools, identity and model.
Evidence
Privacy-safe decision records, dashboards and live policy tracing.
Deployment
EU cloud and on-premise.
-
Next
Policies in natural language
Describe a rule in plain words; Gate writes it as a readable policy you review, and explains every decision in plain words too.
-
Then
Masking and pseudonymisation
Replace personal data with placeholders before the prompt leaves, and keep the rest of the request working.
-
After that
Image content inspection
Apply your rules to images in prompts and answers, not only to text.
-
Later
A web studio for domain specialists
Privacy, legal and business teams write, test and review policies in a web interface, with no code to learn.
Designed and built in Europe, on open and inspectable foundations.
Noozle is an Italian company, and your data stays under EU law, wherever your teams work. The team that wrote the policy language, the matching engine and the gateway integration is the team you talk to. Need a new kind of check? Tell us: we extend the product around what customers need.
Made
and hosted
in Europe
- Tested against the real gatewayend-to-end test suites against a pinned LiteLLM release, deployment canaries and drift checks.
Noozle Streaming
Route the events your business needs.
Gate checks your AI requests and responses. Streaming routes business events to the right destinations. Use the same policy language for both.
Follow one event from its source to every team and system that needs it.
See how Streaming routes events- Kafka
- AMQP
- Pulsar
- NATS JetStream
- Google Pub/Sub
- AWS SQS
- AWS SNS
- AWS S3
- SQL database
- Elasticsearch
- HTTP webhooks
- HTTPS API
- HTTP streams
- SSE
- Push alerts
Articles
Guides, examples, and notes from Noozle.
- LATEST
Your AI Gateway Approved the Request. Should Your Business Have?
Why controlling access to AI models is only the beginning of enterprise AI governance.
Every Prompt Is a Network Call: Why LLM Deployers Need an AI Gateway
Why LLM deployers need an AI gateway, and who should write the rules it enforces.
Your first policies, running in our EU cloud.
Tell us what you want to control. We reply within 24 hours, usually sooner.
Step 1: Assess · next
We map your LLM data flows and privacy risks with you.
Step 2: Go live
Your first policies run in our EU cloud. Your privacy officer reviews the decisions.
Step 3: Choose
Stay in the cloud, or move Gate into your own infrastructure.
Start with one email.
Tell us which gateway you use and what you need to control. We reply within 24 hours, usually sooner.
We use your details only to reply to your request. Privacy policy.
Contacts. Talk to us about Noozle.
Evaluating Noozle Gate or Noozle Streaming? Tell us which product you're interested in and what you need it to do. We're happy to answer questions about capabilities, integrations and deployment options.
Message received. Thank you for contacting Noozle. We’ll reply to your email address.