Giselle Evita Koch

AI Security Services

I audit what your AI agents are actually allowed to do

Prompt filtering changes how often a model is fooled; it does not change what happens when it is. I review the tool-call boundary itself — the last control before an agent's action becomes a side effect — and tell you honestly what I find.

Built on the same work behind agent-security-gate, benchmarked externally on AgentDojo.

Services

Fixed scope, written findings, no retainer required to start.

The evidence behind it

Public repositories, not claims. Each one is reviewable before you hire.

  1. agent-security-gate BSc thesis, TU Darmstadt — policy enforcement at the tool-call boundary, benchmarked externally on AgentDojo's Banking suite
  2. vendor-red-team-passport LLM red-teaming mapped to the OWASP GenAI Top 10, with sanitized evidence and signed release provenance
  3. evidentia Turns IT and security activity into audit-ready evidence — the compliance-evidence reference implementation behind the EU AI Act offer
  4. Current Reviewing the delegation and attenuation invariants of an IETF Internet-Draft on agent authorization, in collaboration with the Tenuo project

Full project writeups and benchmark numbers →

Engagement

Currently taking a small number of pilot engagements.

Pilot pricing

The first engagement is free, in exchange for a public case study or testimonial. Pricing afterward is fixed-scope, quoted directly once we've scoped the work together.

Scope

One agent or AI feature at a time, with a defined set of tools or claims. Clear boundaries, no open-ended engagements.

Turnaround

3–5 business days from kickoff to written report for a single-surface audit.

Get in touch

If a free pilot audit or an EU AI Act gap assessment would be useful for something you're building, I'd like to hear about it.

giselle.evita@gmail.com