Services / 03

AI security assessments

“We’re adding AI to the product.”

I test whether your AI features can expose data or take actions they shouldn’t. That covers the tools, data access, permissions, model loading and deployment pipelines around the model, and the classic application it sits in. You get a report that shows what I found, how to reproduce it and what to fix.

Prompt injection · Agents & MCP · RAG · Model loading · MLOps pipelines

Who it’s for

  • Product teams adding LLM features A chat assistant, a summarizer, a search over your own documents. I check what a user, or a document a user uploads, can make it do.
  • Companies rolling out internal assistants or agents Assistants connected to your tools and data. I test what they can be made to do with the permissions they have.
  • AI vendors You ship models, a platform or an agent framework to customers, and want it tested before they deploy it.
  • Consultancies I also take subcontracting engagements, and I can deliver under your brand with a white-label report.

What I test

I test the system around the model: what it’s connected to, and what it can be made to do.

  • Prompt injection Direct, and indirect through documents, emails and web pages the model reads.
  • Agents, tools and MCP servers What an agent can be made to do with its permissions.
  • RAG and data access Leaks across users, or beyond what the user is allowed to see.
  • Model loading and supply chain Malicious model files, unsafe deserialization, untrusted sources.
  • MLOps pipelines Training, model registry and deployment.
  • The application around the feature The same web, API and access-control testing I do in a classic pentest.

Frameworks: OWASP Top 10 for LLM Applications · MITRE ATLAS · OWASP guidance on agentic AI threats

Certification: OSAI+

How it runs

  1. Scoping & authorization

    Which models, agents, tools and data stores are in scope, which environment I test, dates and limits. I prefer staging; production is tested only within agreed limits, with no destructive tests and in agreed time windows. We also agree whether AI tools may be used on your data: local models by default, cloud models only with your written agreement. Everything is written into signed rules of engagement.

  2. Testing

    Manual, with tooling as support. I record the exact prompts and outputs as I go. Critical findings are reported as soon as I find them, not at the end.

  3. Report & debrief

    Proof, impact and concrete fixes ranked by risk, then a call with your team.

  4. Retest

    Once the fixes ship, I verify them and send an updated report. Evidence and your data stay encrypted during the engagement and are deleted 90 days after the final report.

What you get

  • A report Each finding with its impact and a concrete fix, ranked by risk. In French or English, as you prefer.
  • Transcripts The exact prompts and outputs that reproduce each finding, so your team can replay them.
  • An agent permission review What each agent and tool can access, and what it should be limited to.
  • A debrief call and a retest I walk your team through the findings, then verify the fixes once they ship.
AI & application security / Public sector From an AI chatbot to remote code execution I reached internal services without an account and found a flaw that allowed code execution on a local replica—not the live system. Read the case study

Questions

Do you test the model or the system around it?

The system around it. I test how the model is connected to your tools, data and users, and the application it runs in. I look at model behavior, such as jailbreaks and off-policy output, only when it has a security impact: access to data, or actions taken. Content safety isn’t what this service covers.

Which providers do you work with?

Any. OpenAI, Azure OpenAI, Anthropic, Mistral and self-hosted models are examples, not a limit. What I test is how your application uses the model.

What access do you need?

We agree it at scoping. It depends on what’s in scope: which interfaces, accounts, tools and data stores I can reach. For the agent permission review, I need to know what each agent and tool is connected to.

Does this cover the EU AI Act?

No. A security test can support your compliance work, but it isn’t a compliance assessment under the AI Act.

Talk about your project

I reply within two working days. Tell me what the AI feature does, which models and tools it uses, what data it can reach, and roughly when you want it tested. I send a quote after scoping.