What AI agents can do, where they break and what risks they bring into a company, security included. Without the hype that usually comes with them.
An agent is not a chatbot with a new name: it is a system that receives a goal, picks the steps itself and uses real tools, which means it can get them wrong. That is what makes the topic interesting and risky at once: the same autonomy that saves hours is the one that, without limits, writes in the wrong place.
Here you will find what they can do today, the typical ways they break (hostile instructions hidden in data, permissions that are too broad, no record of what they did) and why a significant share of agent projects gets cancelled before reaching production.
4 articles in this topic, updated 20 May 2026
Information systems are changing: not just interfaces, but APIs, MCP servers and AI agents able to work on company data in a supervised way.
TL;DR - DeepEval is an open source framework that uses a "judge" LLM to evaluate the responses of another LLM, like a professor grading a student's e…
I tested Moltbot, the autonomous AI agent that takes control of your PC. I installed it, tested it and uninstalled it immediately. Here's why autonomy…
Elon Musk imagines a future without traditional applications and operating systems, replaced by AI agents that orchestrate everything. MacroHard repre…
A chatbot answers, an agent acts: it reads systems, takes intermediate decisions and writes somewhere. In practice that means an agent needs tight permissions, limits, and a log of what it did.
They become safe with minimal permissions, an isolated environment and a person approving the operations that matter. The most underestimated risk is prompt injection: hostile content arriving through data and executed as if it were an instruction.
Because they start from enthusiasm for the technology rather than from the cost of the work they are meant to replace. Without that number the project cannot be defended, and it dies at the first cost review.
Take the free assessment: 5 minutes, 5 areas analyzed, personalized PDF report with concrete recommendations.
Find out how AI-ready you areFree, no signup required