Articles from the team
- Microsoft 3657 min read
Microsoft sells five things called Copilot. Here is which one you need.
They are not tiers of one product. They have different prices, different data protections and different jobs — and the confusion is expensive.
- AI Security7 min read
AI did not create new attack techniques. It removed the cost
Targeted phishing used to take hours of research per target. That is the change, and it moves volume and targeting rather than method — which tells you exactly which of your controls degrade and which hold.
- AI6 min read
Governing Copilot Studio: the risk is in the connectors
Copilot Studio comes with your licences, and your teams will build agents with or without a process. Most of the risk sits in what an agent can connect to. What to settle before the first one ships.
- AI5 min read
Your first operations agent: read-only, grounded, and measured
Engineers lose hours rebuilding context before they can fix anything. An agent can gather it, if it reads your real documentation, cannot change anything, and you measure the hours it saves.
- Microsoft 3658 min read
You bought Copilot for everyone. About a third of them opened it.
Microsoft's own adoption numbers are the strongest argument against licensing Copilot by headcount. Here is how to find out who is actually using it, and what to do with the rest.
- AI Security8 min read
The permission you granted an agent is not the permission it has
You gave it read access to one mailbox. It also has every other connected tool, for the whole session. That is the default in most agent frameworks, and it is large by design.
- Microsoft 3659 min read
Fix SharePoint oversharing before you turn Copilot on
Copilot does not bypass your permissions. It reflects them — which is why the first week of a rollout surfaces files nobody meant to share. The cleanup runs on a tool you already own.
- AI Security9 min read
The AI security checklist we actually use, and how to verify each control
Seven sections, each with the check that tells you whether the control exists. Inventory first, because the other six assume you know what is running.
For technical teams
Deeper pieces for developers and security teams: frameworks, protocols and the controls behind them.
- Microsoft 3658 min read
Microsoft keeps your data available. Recovering it is your job.
Retention is not backup, replication is not backup, and the recycle bin empties on a schedule most people have never checked. What Microsoft actually commits to, and what you have to do yourself.
- AI6 min read
AI triage of security findings only works if you ground it
Scanners produce more findings than any team can read, and a model given only the scanner output just guesses faster. What it needs to see, and the controls that make its verdicts worth trusting.
- Compliance9 min read
Mapping one AI agent to three frameworks, and where they run out
An internal IT ticket triage agent, mapped against the EU AI Act, NIST AI RMF and ISO 42001. Three gaps are real, and the security spot-check at the end is the uncomfortable part.
- AI Security8 min read
Three incidents that built the OWASP Agentic Top 10
EchoLeak, Amazon Q and Replit. All documented, dated and public. Two of the three required no attacker sophistication, and one had no attacker at all.
- AI Security8 min read
Getting an AI coding assistant approved instead of blocked
Security says no, the business rolls it out anyway, and now nobody is watching it. The control set that gets a coding assistant approved on purpose.
- AI Security10 min read
OWASP's LLM Top 10 for 2026 is the first one backed by incident data. Three entries moved because of it
The 2026 list weighs practitioner consensus against what has actually gone wrong in the field. Where those two disagree is where your roadmap is probably wrong.
- AI Security11 min read
OWASP has published a Top 10 for agentic AI. If you are running agents, here is what it means for you.
We read all ten entries of OWASP's new Top 10 for Agentic Applications and mapped them to the controls you should already be building — in the order that removes the most risk.
Have a question we have not written about?
Ask us directly. We reply within one business day.