A plain-English field guide to AI security
The Last
Judgment
AI can now write, talk, and act on your behalf. This is the manual for the part that matters: what breaks, who it breaks for, and what fixes it.
Pick the shoes you're standing in
These guides are the same twenty risks, seen from three different chairs. Choose the one closest to yours and the reading order rearranges itself.
AI That Talks
Chatbots and assistants. How they are talked into breaking their own rules, what they leak when they do, and why "it said so" is never a source of truth.
Read the guide → Visited 02 OWASP Top 10 for Agentic Applications · 2026AI That Acts
AI with tools, memory, and permission to press send. Where autonomy stops being useful, and the exact places to put a human instead.
Read the guide → Visited 03 Deepfake incident response guideWhen Eyes Fail
A cloned voice costs seconds to make and has moved $25 million in one call. What to do instead of looking harder: the code word, the callback, the process.
Read the guide → Visited 04 GenAI Red Teaming Guide · 2025How Attackers Think
Nine ways AI speeds up an attacker, and twelve tests you can run against your own agent this week. Plan one in the interactive planner.
Read the guide → Visited 05 GenAI COMPASS Playbook · 2025Your Game Plan
A four-step loop - observe, orient, decide, act - for a threat that keeps changing shape. Includes a severity scorer that turns an argument into a decision.
Read the guide → Visited 06 The fix · Ten habitsHow to Fix It
Every risk in this guide has a control that already works. Here are the ten, with a checklist you can tick off and hand to someone else.
Read the guide →No apocalypse required
AI is not a villain. It is a tool, and like every powerful tool it is exactly as safe as the person holding it.
What this is
A jargon-free translation of the published AI security guidance, with the incident reports that prove each point. Twenty risks, each with a plain sentence of what it is, a real example, and a fix you can actually do.
It is written for the person who has to decide, not the person who has to deploy. That person is increasingly everyone.
What this isn't
Not a prediction about machine consciousness. Not a compliance checklist, and not legal advice. Nothing here asks you to be afraid of AI - the incidents in these pages are mundane, expensive, and almost always preventable with process.
Every problem has already happened to somebody. Every one of them already has a fix.
Talks versus acts
Nearly every AI headline makes sense once you know which of these two you are looking at. One writes things. The other does them. Only one of them can wire money at 3am.
It produces text
A large language model predicts the next token. Its output is a suggestion until something downstream trusts it. Most of its risk is inherited: it becomes dangerous when its words are executed, rendered, or believed.
It takes actions
An agent is a model in a loop with tools and memory. It plans, acts, reads the result, and repeats. That loop is what makes it useful - and what turns one bad instruction into a sequence of real-world effects before anyone reviews them.
Myth, meet reality
Tap any card to see what the evidence actually says. All four beliefs are reasonable - three of them are also the reason incidents keep happening.
How exposed are you, really?
Six questions, no sign-up, nothing sent anywhere. Answer honestly - the result is a reading order, not a grade.
The 60-second self-check
0 of 6 answeredIf your boss appeared on a video call and asked you to move money today, would you verify it on a number you already have?
Could you name every AI tool your team uses with company data?
Does something an AI drafts get reviewed by a person before it is sent or published?
Do you know who can switch off the AI features your work depends on?
Do you have a code word or callback rule that a cloned voice could not get around?
Do your AI tools have a spending limit and a rate limit you could point to?
The clock is friendly
This is not a doomsday timer. It is a reminder that the cheapest moment to learn good habits is before you need them.
Plenty of time. No panic required - just a little care, applied early.
The ink isn't dry yet
Nothing in these pages is inevitable. The future of AI will be decided by ordinary choices: whether we read the manual before handing over the keys, whether someone is allowed to ask "are you sure", whether the code word actually gets used.
That is all it takes to turn a scary story into a boring one - and boring is the goal.
A product of Kulutchka
This guide was researched, written, designed and built by Kulutchka International, an AI-security consultancy that helps boards and executive teams understand, govern and defend the AI their organisation already uses.
The register this site is modelled on is published at kulutchka.com: forty-six defined services across twelve practice areas, from strategy and governance to red teaming and incident response.
- Produced by
-
Kulutchka International
kulutchka.com - Illustrations
- Original isometric plates, generated with AI and art-directed to this guide's palette. No stock art, no borrowed illustrations.
- Type
- Anton, Space Grotesk and JetBrains Mono, served by Google Fonts.
- Sources
- The OWASP GenAI Security Project: Top 10 for LLM Applications (2025), Top 10 for Agentic Applications (2026), the GenAI Red Teaming Guide and the COMPASS playbook, all CC BY-SA 4.0. This is an independent summary for general readers, not an OWASP publication.
- Rights
- © 2026 Kulutchka International. Text and illustrations are CC BY-SA 4.0, matching the licence of the source material.
- Privacy
- No analytics, no cookies, no accounts. Answers you give in the interactive sections stay in your browser.