Test The Model, Not Just The Application Around It

Your AI is in production, and the penetration test you ran covers the web application in front of it, not the model behind it. IrisInfosec closes that gap. LLM and agent red teaming against the OWASP Top 10 for LLM Applications, adversarial testing for poisoning, evasion and extraction, and security across the AI supply chain.

10

Risk categories in the OWASP Top 10 for LLM Applications, tested end to end.

ATLAS

Findings mapped to MITRE ATLAS adversarial techniques.

3

Layers tested: the model, the pipeline that ships it and the agent that acts on it.

Article 15

AI Act accuracy, robustness and cybersecurity duties evidenced.

What this covers

The attack surface your application test does not reach

A standard application test looks for injection in the input field, flaws in the session and gaps in the API. None of that finds a document in your knowledge base that carries instructions for the model, an agent with a token that can do more than its task requires, or a fine-tuning set that someone else contributed to.

Those are model-layer problems, and they behave differently from software bugs. There is often no patch, only a control. That is what we test for, and what we help you design around.

In one line

AI security is the practice of testing and defending machine learning systems themselves: the model, the data that trains it, the pipeline that ships it and the agent that acts on its output. It is distinct from AI governance, which sets the policy, classification and accountability around those systems.

What we test

Where generative systems actually fail

Prompt Injection, Direct And Indirect

Instructions delivered through a user message, a retrieved document, a web page or a file the model reads. Indirect injection is the one most teams have never tested for.

Data And Model Poisoning

Manipulation of training, fine-tuning or retrieval data to change behaviour on a chosen trigger, including contributions that arrive through the supply chain.

Model Extraction And Inversion

Systematic querying to reconstruct model behaviour or recover training data, including membership inference against sensitive records.

RAG And Vector Store Weaknesses

Embedding collisions, cross-tenant retrieval, poisoned chunks and access controls that the retrieval layer quietly bypasses.

Excessive Agency In Agents

Tool permissions, credential scope, chained actions and the human checkpoints that should exist between a model's suggestion and a real world effect.

AI Supply Chain And MLOps

Model provenance and signing, notebook and pipeline security, registry and artefact integrity, and the third-party components a model quietly depends on.

Common question

AI security and AI governance are not the same work

Buyers often ask for one and need both. They answer different questions, use different methods and usually belong to different owners.

AI security(this page) AI governance
Question answered Can this system be manipulated, extracted or poisoned Are we permitted to run this system, and who is accountable
Method Red teaming, adversarial testing, pipeline and agent review Inventory, risk classification, management system, oversight design
Reference OWASP LLM Top 10, MITRE ATLAS, NIST AI 100-2 EU AI Act, ISO 42001, NIST AI RMF
Output Reproducible findings and a hardening plan Technical documentation, oversight controls and an audit trail
Usual owner Security engineering and the product team Legal, risk and the AI governance function

How we work

Automated coverage first, then the attacks a person has to think of

Automated sweeps establish the baseline quickly and cheaply. The findings that matter usually come from a human working the system's specific context.

Context decides severity. An attack path in one deployment is harmless in another, so the threat model comes before the test plan.

Proof is the same attack, run again. Guardrails move behaviour rather than removing a defect, so a control is only proven when the attack that beat it fails.

Stage What we do
Threat model first Map the RAG pipeline and multi-agent workflow, including data flows, retrieval sources, tools, credentials and trust boundaries, before any testing
Automated sweep Broad coverage across injection, jailbreak, extraction and output handling to clear the known ground fast
Manual red teaming Context-specific attacks built from your business logic, your knowledge base and the actions your agents are allowed to take
Retest after hardening Re-run the successful attacks against the hardened system to confirm each control holds

Process

From scope to a model you can defend

A measured five-stage approach that ends in findings your engineers can reproduce.

01

Scope And Threat Model

Map models, data flows, tools, permissions and trust boundaries, including the AI in vendor features nobody registered.

02

Automated Sweep

Broad adversarial coverage across the OWASP LLM Top 10 to establish the baseline.

03

Manual Red Teaming

Context-specific attacks against your prompts, your retrieval layer and your agent permissions.

04

Harden

Guardrails, output handling, permission scoping and pipeline controls designed to your risk, not to a template.

05

Retest And Monitor

Re-run the successful attacks, then set the monitoring and re-test cadence for model and prompt changes.

What you receive

Findings an engineer can reproduce and a regulator can read

AI system inventory and threat model

Every model in production, built and bought, with data flows, tools and permissions mapped to MITRE ATLAS.

LLM red team report

Successful attacks with the exact prompts, payloads and conditions, so any finding can be reproduced.

Adversarial robustness assessment

Poisoning, evasion, extraction and inversion results against your model and your data.

AI supply chain and MLOps review

Provenance, signing, registry integrity and pipeline access, with the weak links named.

Guardrail and control design

Input handling, output validation, permission scoping and human checkpoints specified for your architecture.

Evidence pack for the AI Act

Testing records structured against the accuracy, robustness and cybersecurity duty in Article 15.

Find out what your model does under pressure

Ask For A Scoped LLM Red Team

Tell us what the system does, what it can reach and what it is allowed to act on. We will
send a scope and a test plan, usually within 3 business days.

Questions, answered

AI Security and GenAI Protection FAQ's

AI security tests and defends the system: the model, the training and retrieval data, the pipeline and the agent. It asks whether the system can be manipulated, poisoned or made to leak. AI governance sets the policy around it: inventory, risk classification, accountability and oversight, aligned to the EU AI Act, ISO 42001 and the NIST AI Risk Management Framework. Most organisations need both, and they are usually owned by different teams.

Prompt injection is the delivery of instructions to a model through content it processes, either directly from a user or indirectly through a document, a web page or a file it retrieves. It is not a defect in a specific line of code. It follows from the fact that a language model does not reliably separate instructions from data. That is why it is managed with controls, isolation and least privilege rather than closed with a patch, and why it needs retesting after every material change.

Both, and the application test alone is what leaves most organisations exposed. We test the interface and the API in the usual way, then we test the model layer: injection through retrieved content, poisoning of training and fine-tuning data, extraction and inversion, retrieval controls, and what an agent can actually do with the credentials it holds.

MITRE ATLAS is a knowledge base of adversarial tactics and techniques observed against machine learning systems, structured the way MITRE ATT&CK is structured for enterprise systems. Mapping findings to ATLAS gives your security team one vocabulary across conventional and AI threats, and it makes an AI red team report legible to a SOC that already works in ATT&CK.

For high-risk systems, Article 15 requires an appropriate level of accuracy, robustness and cybersecurity across the lifecycle, including resilience against attempts to alter use or behaviour by exploiting vulnerabilities. It names data poisoning, model poisoning, adversarial examples and confidentiality attacks as risks to be addressed. It does not prescribe a test method, so the practical answer is testing that is documented, repeatable and mapped to a recognised taxonomy.

The Digital Omnibus on AI, in force since July 2026, deferred the application dates for high-risk obligations. Stand-alone high-risk systems under Annex III move to 2 December 2027, and high-risk AI embedded in regulated products moves to 2 August 2028. Obligations for general purpose AI models and the prohibited practices were not deferred. The requirements did not change, the preparation window did.

Yes, and it is where the sharpest findings tend to be. We map each agent's tools, credentials and permitted actions, then test what happens when the model is persuaded to use them: chained actions, privilege inheritance between agents, tool responses treated as trusted input, and the absence of a human checkpoint before an action with real world consequence.

Yes. Most production risk sits in systems assembled from foundation models, vendor features and third-party components rather than models trained in-house. We test the deployment: your prompts, your retrieval layer, your tools, your permissions and your output handling. Where a vendor's own model is in scope, testing stays within their acceptable use terms and we agree the boundary with you first.

Related services

Find The Failure In Your AI Before Someone
Else Does

Book a call with a team that tests models and reads the AI Act at
source. We will send a scope and a staged plan, no obligation.