İçeriğe geç

AI security

AI is opening up a new attack surface.

Prompt injection, overstepping, data leakage — these are not theoretical, they are what we find in established systems. We take into account when installing, and we also search in established systems.

Four headings

During deployment and after go-live.

  • Safe installation

    It's part of every AI system we install, not a separately purchased add-on. The assistant knows what he can see on behalf of whom; he also knows what he can't see. What he does is recorded.

    • Prompt injection defense
    • Authority-based data access
    • Output filtering
    • Audit records
  • AI system audit

    We try to break an already installed AI application: leaking the system instruction, exceeding the authorization limit, moving out data that should not be accessed. We report the findings in order of importance and in reproducible steps.

    • Red-teaming
    • Prompt injection testing
    • Exceedance attempts
    • Priority finding report
  • EU Artificial Intelligence Act — technical side

    We establish the technical mechanisms required by law: record keeping and traceability, human oversight step, system documentation. It is the job of the legal evaluation lawyer; it is our job to leave the infrastructure on which that evaluation is based working.

    • Recording and traceability
    • Human surveillance step
    • System documentation
    • Technical support to risk classification
  • AI in security operations

    Underneath the security team is often a pile of warnings. We automate the repetitive triage and intervention steps, ensuring that the decision remains with the person. We start by keeping the scope narrow and measuring — silent error on the safety side is the most expensive error.

    • Alert prioritization
    • Log analysis
    • Incident response automation
    • Narrow scope pilot

The evidence

We have established systems that work without data ever going out.

Bizmed's clinical platform operates in nearly a hundred hospitals, each institution's own data center. Closed network, on-premise Oracle, no outbound connection. Most of the constraints we talked about in AI security were experienced before AI.

Bizmed

  • on-premise
  • closed network
  • Oracle
  • authorization-based access

Do you have an AI system installed?

Tell us what you connected to where — where you leaked it, what you need to record.