AI security
Everything I've written on ai security, with the research pieces first. Each one ends with something you can do on Monday.
Dealing with one of these problems yourself? I work with organizations that need answers, not another slide deck. Let's talk.
What's in a five-angle short
Most subjects here are one page with five short sections, always in this order. Each has its own colour inside the article so you can jump to the one you need.
- 1Where I'd start: the first move I'd make, and the question I'd ask before spending anything.
- 2Where plans fail: where plans on this subject usually break, and the assumption I'd test first.
- 3Beyond the jargon: what the subject actually means once the vendor language is stripped out.
- 4What good looks like: the result that would tell you it worked, and the number I'd watch.
- 5The call leaders can't dodge: the decision leadership has to own, and what it needs on the table to make it.
-
Before You Give an AI Agent Permission to Act
AI agents can take action, not just answer questions. I examine permissions, tool access, human oversight and what it takes to keep automation accountable.
-
Before an AI Agent Can Act, Decide What It Must Never Touch
What I look at in AI agent permissions, the decisions it forces, and what credible evidence looks like before you spend.
-
How Do You Keep an AI Agent Within a Human Approval?
Bind AI-agent approval to the exact action, target and payload. Handle changed data, delegated credentials, retries and uncertain execution without widening consent.
-
How Do You Stop AI Agents Acting on Instructions in Documents?
Constrain retrieved content with task-bound tool authorization, verified recipients, data-flow controls and approval checks that survive model mistakes.
-
How Do You Stop a RAG Assistant Revealing Documents After Access Is Revoked?
Trace revoked document access through RAG permissions, indexes, caches and conversations. Define a defensible revocation boundary and test it with synthetic data.
-
How Should a Company Approve Employee Use of Generative AI?
Approve employee AI use by task, data and account—not brand alone. Assess contracts, connectors, review capacity and conditions for withdrawing permission.
-
AI Governance: Who Can Actually Stop the System?
A policy saying humans remain in control is easy to write. I want to know what happens when an AI system is producing results everyone likes, the deadline is...
-
AI agent permissions
I would not begin AI agent permissions by buying another platform or launching a large programme. I would find the boundary, the accountable owner, and one s...
-
AI governance
I would not begin AI governance by buying another platform or launching a large programme. I would find the boundary, the accountable owner, and one scenario...
-
AI phishing
I would not begin AI phishing by buying another platform or launching a large programme. I would find the boundary, the accountable owner, and one scenario w...
-
AI risk communication
I would not begin AI risk communication by buying another platform or launching a large programme. I would find the boundary, the accountable owner, and one ...
-
An AI Agent With Permission to Act Is a Security Architecture Problem
An AI assistant that answers a question is one kind of risk. An AI agent that can read a customer database, send messages, change permissions, and execute wo...
-
Before You Give an AI Agent Permission to Act
An AI system that answers a question can mislead you. An AI system that can open a ticket, change a configuration, retrieve private records, or send a messag...
-
Can You Control AI Without Understanding Every Answer?
Learn how AI control can focus on authority, evidence and consequences without requiring complete understanding of every internal computation.
-
LLM data leakage
I would not begin LLM data leakage by buying another platform or launching a large programme. I would find the boundary, the accountable owner, and one scena...
-
Model evaluation
I would not begin model evaluation by buying another platform or launching a large programme. I would find the boundary, the accountable owner, and one scena...
-
Model monitoring
I would not begin model monitoring by buying another platform or launching a large programme. I would find the boundary, the accountable owner, and one scena...
-
Model supply chain
I would not begin model supply chain by buying another platform or launching a large programme. I would find the boundary, the accountable owner, and one sce...
-
Prompt injection
I would not begin prompt injection by buying another platform or launching a large programme. I would find the boundary, the accountable owner, and one scena...
-
RAG security
I would not begin RAG security by buying another platform or launching a large programme. I would find the boundary, the accountable owner, and one scenario ...
-
What to ask an AI vendor before your data goes in
Procurement questionnaires for AI products tend to copy the ones for SaaS, and they miss the questions that matter.
-
When Does an AI Warning Become a Working Control?
Turn AI risk warnings into owned, funded and testable controls. A short from Boxing Einstein on the gap between recognizing a problem and changing it.
-
Why Useful AI Can Be Harder to Govern
Why a useful AI answer still needs evidence before wider use: distinguish observed benefits, hidden consequences and the cost of waiting.
-
Your staff are already using AI. Now what?
Banning it doesn't work. I've yet to see an organisation where a policy stopped people pasting a contract into a chatbot to summarise it. What the ban does i...
Nothing matches. Try another word, or ask me directly.