AI Security Review
A 30-minute technical conversation about your AI deployment and the attack surface it creates.
For security teams and engineering leaders running LLM and agentic systems in production. No charge, no sales pitch.
The gap
Most organizations find their AI security problems after deployment. The questions that should have been settled during architecture:
- Where can a prompt injection reach in your agent’s tool chain?
- Can a poisoned document in your RAG pipeline change model behavior?
- Are tenant data and context actually isolated in your vector database?
- What can your agents do, and what stops them from doing it by accident?
- How do you reconstruct an LLM decision when an auditor asks?
- Which of your MCP servers exposes an internal system through an AI interface?
SOC 2, ISO 27001, and your Kubernetes policies weren’t written with any of these in mind. Neither was your standard vendor assessment.
What 30 minutes actually gets you
I want to be precise about the scope, because half an hour is half an hour.
A map of your top attack paths. Walking your architecture together, I can usually identify the three or four places where your specific setup is exposed — prompt injection reach, agent privilege scope, data leakage vectors, tenant isolation.
A prioritized shortlist. What to fix before you ship, what can wait a sprint, what is genuinely low risk. Ranked, with reasoning.
Pointers to working patterns. Where I’ve seen an architecture like yours solved before, I’ll say so and point you at the reference implementation or the writeup.
What this is not: a documented threat model, a written report, or a full architecture review. Those take days, not minutes. If you need one, we can talk about it separately.
Who this is for
- Security teams evaluating an AI system before production, or responding to a review request
- CTOs and engineering leaders who need to understand what risk their AI deployment introduces
- Platform teams building Kubernetes-based AI/ML platforms who need to scope LLM workloads correctly
- Compliance and risk officers extending an existing program (SOC 2, ISO 27001, GDPR, HIPAA) to cover LLM and agentic systems
How the call runs
Current state (10 min) — your architecture: models, agents, data flows, integrations. What your agents can reach. Which existing controls apply to the AI stack and which quietly don’t. Compliance constraints.
Risk identification (15 min) — prompt injection and indirect injection paths, data privacy across training data and RAG context and outputs, agent privilege scope and sandboxing, multi-tenant isolation at the model and vector database and infrastructure layers, audit and observability gaps.
Action plan (5 min) — highest-priority findings with specific remediation approaches, quick wins versus architectural changes, what to bring back to your team.
Background
Amine Raji, PhD, CISSP — AI security researcher, with 15+ years securing production systems in banking (Société Générale), defense, aerospace (Airbus), and automotive (Volvo Cars).
My work is on AI systems once they are deployed and running: what an agent can reach in a live environment, how far an injected instruction travels through it, and which controls hold up under load and under attack. I run the attacks in a lab before I describe them, which is why the findings come with reproductions rather than assertions.
Work you can check before you book
I publish the research rather than asking you to take it on faith:
- MCP Security, One Year In — OWASP Stockholm, May 2026. Three MCP attack patterns (tool poisoning, meta-context injection, cross-server hijacking) with live demonstrations. Slides.
- The AI Agent Security Field Guide — the OWASP Agentic Top 10 mapped to real attack patterns and specific controls.
- mcp-attack-labs — reproducible labs for the attack patterns above. Run them yourself.
- procurement-ai — a reference deployment showing the controls in place from the first commit rather than retrofitted.
- Hidden risk in MCP servers and the rest of the blog — production architectures, attack scenarios, working code.
- Contributor to the OWASP Agentic Security project. Peer-reviewed publications in cybersecurity and distributed systems.
Every claim I make in front of an audience has been reproduced in a lab first. The same standard applies on this call.
Book a slot
30-Minute AI Security Review
Prefer email? [email protected]
All conversations are confidential. NDA available on request.
Questions
Why is there no charge? The industry is repeating the “ship first, secure later” mistake it made with web applications twenty years ago, and I’d rather see the problems early than write about them afterward. These calls also feed my research — the patterns I see across deployments are what I end up publishing. That’s the trade.
Will you try to sell me something? No. I’ll tell you what I find, including when your architecture is already sound. If a deeper engagement makes sense we can discuss it afterward, but the call stands on its own.
We’re a security team. Won’t we already know this? Sometimes. But prompt injection, indirect injection through RAG, agent tool misuse, and vector database isolation don’t map cleanly onto existing threat frameworks, and most enterprise security programs weren’t built to evaluate them. A second read from someone who builds these systems is usually worth half an hour.
Our system is already in production. Too late? No — it’s useful timing. I can map the attack surface against your live architecture and prioritize by what presents immediate risk rather than theoretical risk.
We’re still evaluating vendors. Too early? Better, actually. Security designed into an architecture costs a fraction of retrofitting it. I can help you ask vendors the right questions and avoid decisions you’d have to undo.
What should we prepare? A working understanding of your architecture: which models or LLM services you use, what your agents can do, what data flows through the system, and where it runs. I’ll ask the rest on the call.
We’re under NDA. Standard for this work. We can sign one beforehand. I work regularly with defense, banking, and automotive clients under strict confidentiality.
Want AI security research in your inbox? Subscribe to the newsletter for analysis of LLM vulnerabilities, agentic system risks, and what security teams need to know about AI deployment.