DField SolutionsMérnöki stúdió · Budapest
Loading · Töltődik
Skip to content
CASE STUDIES · saas

Hardening AI agents with Noaber AI

With Joran van Beek of Noaber AI, we studied how autonomous AI agents get attacked once they run continuously with real capabilities — and what a trustworthy "hardened shell" around them looks like. The work became a co-authored, openly published paper.

TimelineCo-authored research
← Back to case studies
Reviewed by
01Co-authored, openly licensed research — "The Hardened Shell" (CC BY 4.0, DOI 10.5281/zenodo.18471237).
02A reusable framework for evaluating agent safety and data sovereignty before shipping.
03Offensive-security depth applied directly to AI agents — the same rigour behind our client work, made public and citable.
The problem01 / 03
  • 01Autonomous agents now run continuously on a user's behalf — reading private data, executing shell commands and retaining memory — using capabilities originally designed for isolated contexts.
  • 02Assembled from isolated components, these agents inherit outdated trust assumptions about locality, memory integrity and tool trustworthiness that existing security models don't address.
  • 03This "Agentic Paradox" opens whole new classes of manipulation — prompt injection and beyond — that standard application-security thinking misses.
  • 04Widely-adopted agent designs (the OpenClaw pattern) shipped with little public scrutiny of their security architecture.
The solution02 / 03
  • 01Red-teamed the OpenClaw agent architecture to map the new manipulation classes that autonomy exposes.
  • 02Reframed the problem around trust boundaries, identity and execution constraints — a "hardened shell" model for safety and sovereignty.
  • 03Evaluated instruction-following and prompt-injection resistance instead of trusting demo behaviour.
  • 04Published the analysis and methodology openly so other teams can reuse it.
The outcome03 / 03
  • 01Co-authored, openly licensed research — "The Hardened Shell" (CC BY 4.0, DOI 10.5281/zenodo.18471237).
  • 02A reusable framework for evaluating agent safety and data sovereignty before shipping.
  • 03Offensive-security depth applied directly to AI agents — the same rigour behind our client work, made public and citable.
CASE STUDIES

Let's get started.

Send an email or book a 30-minute call.

Back to case studies