AI & LLM security

AI and LLM security audit for chatbots, assistants and agents.

If you have put an AI assistant in front of customers or connected one to internal data, attackers can talk to it too. I test what it can be made to say, reveal and do.

Certified CLLMSP (LLM security), with testing based on the OWASP Top 10 for LLM Applications.

Book a free 30-min callWhat I test
At a glance
Targets
Chatbots · RAG · AI agents
Reference
OWASP Top 10 for LLM Applications
Compliance
EU AI Act readiness
Deliverable
Report, debrief and retest
01 / Scope

What I test

The model is only one part of the attack surface.

Prompt injection

Direct and indirect: users, documents, emails or web pages that hijack the instructions your assistant follows.

Data leakage

System prompts, other users’ conversations, internal documents pulled in by retrieval, secrets in context.

Unsafe actions

Tools the agent can call, such as emails, refunds or database queries, and what happens when it is tricked.

The application around it

API, authentication, rate limits and logging around the model, where classic vulnerabilities still apply.

02 / Method

How an AI security audit runs

Built for teams shipping AI features fast.

  1. Scoping. Use cases, model provider, data sources and tools connected to the assistant.
  2. Threat modelling. Who can talk to it, what it can access, what it can do, and what would hurt your business most.
  3. Adversarial testing. Manual prompt injection and jailbreak campaigns, combined with classic tests of the surrounding application.
  4. Report. Findings ranked by business impact, with practical mitigations: privilege separation, filtering, human approval for risky actions.
  5. Debrief and retest. We review the results together, then I re-test after your changes.
03 / Typical findings

Where AI assistants usually break

Most issues come from what the assistant is allowed to see and do.

System prompt disclosure

Internal instructions, business rules or credentials revealed with a few well-chosen questions.

Indirect injection

A document, email or web page the assistant reads silently takes control of its answers or actions.

Over-privileged tools

The assistant can call functions or data it never needs, so a single injection has a large impact.

No abuse monitoring

Nobody would notice a user systematically extracting data or testing jailbreaks.

04 / FAQ

AI security audit: your questions

Our chatbot uses a provider’s API. Isn’t the provider responsible?

The provider secures its model. How you connect it to your data, tools and users is your responsibility, and that is where most real-world issues come from.

What does EU AI Act readiness mean here?

The AI Act sets robustness and cybersecurity requirements for certain AI systems. The audit documents how your system resists attacks and which gaps to close. It is not legal advice.

Can you test before we launch?

Yes, and it is the best moment: fixing an over-privileged tool or a leaky prompt before go-live costs far less than after an incident.

Do you need access to the model itself?

No. Access to the application as a user, plus documentation of its tools and data sources, is enough. A test environment is ideal.

Shipping an AI feature?

Tell me what your assistant does and what it can access. You get a fixed-price proposal after a free 30-minute call.

Book a free callRead: direct prompt injection