I'm Amine.
I help teams build AI systems 
that don't become security incidents.

For fifteen years I secured things where the consequences were physical. Grid operators, hospital systems, payment infrastructure. You learn to think in blast radius rather than vulnerability counts, because the question anyone senior asks is never "how many findings" but "what happens if this one goes wrong at 3am".

Then I started running AI agents with real credentials across multi-cluster Kubernetes estates, and found the same question had no good answer. An agent reads a document, decides something, and calls a tool holding permissions nobody threat-modelled. None of the scanners I had trusted for a decade looked at that path.

I kept meeting the same standoff. Engineering teams shipping capable AI systems with no security expertise in the room, and security teams asked to sign off on systems they had no way to reason about. Both were right, and neither could close it alone. I do both jobs, so I sit in the middle.

I contribute to the OWASP Top 10 for Agentic Applications 2026 and speak on agent security. The most useful thing I have published recently is the OWASP Stockholm debrief on a year of hands-on MCP work: which attacks kept recurring, and which defences held.

The short version of what I have learned: prompt-level controls fail and infrastructure controls hold. The attacks are genuinely new. The defences that survive contact are scoping, identity and segmentation, which your team already knows how to enforce.