Skip to content
Edictum Labs, home
Discipline 03

AI
security

Agents, inference, and native execution. AI systems now run code, call tools, and reach real resources — which makes them a real security surface, not a conversational one.

Discuss an AI assessment
Scope
Agents · Inference · Runtimes
Lens
Execution & authorization
Based in
Prague
The shift

An AI system that can act is a system that can be made to act.

The risk is not that a model says the wrong thing. It is what the system around it is allowed to do.

Once a model can call tools, run code, or reach data, the security question stops being about language and starts being about capability. An adversarial input — from a user, a retrieved document, or an upstream model — becomes an attempt to drive real actions against real systems.

We assess AI systems the way we assess any other execution surface: by asking what it can actually reach, and proving how far a hostile input gets.

Surfaces

Where AI systems actually expose risk.

The boundaries that matter once inference is wired into tools, data, and infrastructure.
  • AI agents and tool-execution boundaries
  • LLM application and prompt-handling security
  • Inference runtimes and native execution paths
  • Model-serving infrastructure and isolation
  • Authorization between agents, tools, and data
  • Supply chain for models, weights, and extensions
Approach

How we test AI systems.

01

Treat the agent as untrusted

An agent that can call tools is an execution surface. We assess what it can actually reach — the filesystem, the network, internal services — when an instruction, a document, or a model output turns adversarial.
02

Follow the native code

Modern inference runs through native libraries, custom kernels, and model-serving stacks. We look at the memory handling and trust boundaries behind the abstraction, where the real corruption bugs live.
03

Test the authorization, not the prompt

Prompt injection is a symptom. The question that decides impact is what the system is permitted to do once injected — and whether tool, data, and identity boundaries hold under a hostile input.
04

Separate demo from deployment

A model that behaves in a demo can still expose a real system in production. We assess the integration — the glue between inference and the resources it touches — not the model in isolation.
What you receive

A clear picture of real AI risk.

Grounded in demonstrated reach, not speculative prompt tricks.
  • Threat model for the AI system and its tool surface
  • Demonstrated findings with real reach and impact
  • Authorization and isolation boundary analysis
  • Native-execution and runtime review where relevant
  • Remediation guidance for engineers and model owners
  • Retest of fixes against the original and variant cases
AI security

Shipping agents or inference to production?

If your AI system can call tools, execute code, or touch sensitive data, it deserves a real assessment. Tell us how it is wired, and we will tell you where it is exposed.

Reply
Within two business days
Authorization
Written, before any testing
Based in
Prague, Czech Republic