---
title: "AI security"
description: "Security for LLM applications and AI agents: prompt injection, data leakage, agent permissions, red teaming and Admina controls, following OWASP and the AI Act."
url: "https://www.noze.it/en/cybersecurity/ai-security/"
lang: "en"
alternate: "https://www.noze.it/cybersecurity/sicurezza-ai/"
---

# AI security

Prompt injection, agent permissions, data leakage: we secure LLMs and AI agents before and after release.

## New attack surfaces

An application with an LLM reads text it does not control: documents, emails, web pages, answers from other systems. An agent also has tools and permissions to act. That is why AI security does not stop at traditional controls: you need to check what the model can read, what it can do and what gets recorded.

We use the [OWASP Top 10 for LLM Applications 2025](https://genai.owasp.org/llm-top-10/) and the [OWASP Top 10 for Agentic Applications](https://genai.owasp.org/resource/owasp-top-10-for-agentic-applications-for-2026/), published in December 2025, as references.

## The main risks

### Prompt injection

Instructions hidden in documents, emails or web pages that the model reads and mistakes for commands. It is the top risk (LLM01) in the OWASP Top 10 for LLM Applications.

### Data leakage

Personal data, company secrets or the system prompt ending up in answers or in external services.

### Agents with too much power

An agent with broad tools and permissions can carry out actions nobody authorised: the “excessive agency” (LLM06) of the OWASP Top 10 for LLM Applications.

### Poisoned data and models

Manipulated training data or components that alter the model’s behaviour. For high-risk systems, the AI Act also requires measures against these attacks.

### Unchecked outputs

Model-generated text inserted without checks into code, queries or commands, with effects on downstream systems.

### Unbounded consumption

Loops and requests that grow out of control, driving up costs and causing outages.

## What we do

- **Assessment and red teaming** of LLM applications and agents: direct and indirect prompt injection, data leakage, tool abuse.
- **Agent perimeter**: allowed tools, minimum permissions, budgets, exit conditions and human approval of actions with impact.
- **Runtime controls** with [Admina](https://www.noze.it/en/open-source/admina/), the open-source AI governance framework, a project by Stefano Noferi, sponsored by noze: prompt injection filtering, personal data redaction, policies on outbound destinations, sandboxes for tools and an audit trail with a SHA-256 hash chain.
- **Evaluations before release** and monitoring in production.

For high-risk systems, Article 15 of the AI Act requires an appropriate level of accuracy, robustness and cybersecurity, with measures against data and model poisoning, adversarial examples and confidentiality attacks; under the Digital Omnibus these obligations apply from 2 December 2027 for Annex III systems and from 2 August 2028 for those embedded in products. We design the controls following the [OISG](https://www.noze.it/en/insights/oisg-paradigm-autonomous-ai/) paradigm, which assigns a metric to every pillar. For how we build agents, see [Agentic AI](https://www.noze.it/en/artificial-intelligence/agentic-ai/).

## Frequently asked questions

What is prompt injection?

It is an attack in which text read by the model (a document, an email, a web page) contains instructions that the model follows as if they came from the user or the developer. With agents, which have tools and permissions, the consequences can be real actions.

Is an anti-injection filter enough?

No. Filters are heuristic: a useful signal, not a guarantee. That is why we combine them with minimum permissions for every tool, policies on outbound destinations, sandboxes, human approval of actions with impact and an audit trail.

How do you limit what an agent can do?

With a perimeter decided in advance: allowed tools, least-privilege permissions, allowed destinations, budgets and exit conditions for the loop. Actions with impact remain proposals that a person approves.

Do you also test models and agents run locally?

Yes. We assess applications built on commercial models and on open-weight LLMs run on-premise, together with the tools and data they can access.

Contact us

## Want to put your AI application to the test?

Fill in the form: your request goes to the dedicated team.

[Contact us now →](https://www.noze.it/en/contact/?topic=cyber)
