AI
Everything I’ve written on AI, newest first.
New
- AI · 23 October 2026
LiteLLM hit again: protect your release pipeline
A malicious LiteLLM release was live for about two hours, published through the project’s own workflow with stolen engineer tokens.
- AI · 8 October 2026
Know what your AI agents did
When an AI agent sends an email or changes a record, you need to be able to see it later. Logging is the boring part that matters.
- AI · 17 September 2026
Six questions before you buy an AI tool
Every vendor now has AI features. A few direct questions tell you what happens to your data.
- AI · 25 August 2026
AI browsers act with your logged-in sessions
AI browsers and browser agents click around for you, using every site you’re signed into. A malicious page can talk to them.
- AI · 9 July 2026
AI coding agents can read your .env file
An agent working in your project folder sees the same files you do, including secrets. Keep real keys out of its reach.
- AI · 18 June 2026
Local AI models for sensitive work
Open models now run well on a good laptop or a small server. For some confidential tasks, that’s simpler than any contract.
- AI · 19 May 2026
The EU AI Act delay: what moves and what doesn’t
EU lawmakers agreed to push back the high-risk AI rules. Transparency duties and the bans stay. And it isn’t law yet.
- AI · 12 May 2026
LiteLLM proxy SQL injection: patch your AI gateway
A critical bug in LiteLLM’s proxy lets attackers read its database without logging in. Exploitation started within days.
- AI · 30 April 2026
An AI usage policy that fits on one page
Long AI policies don’t get read. One page with clear yes and no answers does more.
- AI · 2 April 2026
AI invents package names. Attackers register them
Coding assistants sometimes invent package names. Attackers publish malware under those names and wait.
- AI · 27 March 2026
Backdoored LiteLLM: your AI gateway holds every key
Two LiteLLM releases on PyPI stole credentials, using a token taken in the Trivy compromise. AI tooling is now supply-chain target number one.
- AI · 3 February 2026
Self-hosted AI agents, open to the internet
Tens of thousands of OpenClaw agents were found reachable online, many unpatched. An agent with your keys needs a locked door.
- AI · 27 January 2026
Shadow AI: find the tools people already use
Your staff already use AI tools you never approved. Banning them doesn’t work. Find them, then offer a safe option.
- AI · 22 January 2026
Before you connect an MCP server
MCP servers let AI assistants use your tools. Each one is code you run with your permissions. Check it like any other software.
- AI · 15 January 2026
Give AI agents the least access that works
An AI agent with access to your mail, files and admin tools can be talked into using all of them. Start small.
- AI · 18 November 2025
Before you ship code an AI wrote
AI coding tools are fast and often right. Five checks before that code goes live.
- AI · 12 August 2025
Prompt injection, explained for decision makers
The text an AI assistant reads can give it orders. You can’t patch it away, so limit what it can do.
- AI · 11 February 2025
Using AI at work without leaking company data
Your team already pastes work into AI chatbots. A five-line policy that keeps it safe.
- AI · 6 August 2024
When the boss on the video call isn’t real
An employee sent $25 million after a video call where everyone else was a deepfake. The fix is a process, not a tool.
- AI · 5 March 2024
Before you turn on Copilot, fix your sharing
An AI assistant sees everything the user can see. Your old oversharing becomes very easy to find.
- AI · 21 November 2023
Before you upload company files to a custom GPT
Custom GPTs are easy to build. Their knowledge files are also easy to pull out. Assume users can read everything.
- AI · 15 August 2023
Read the AI clause in your software terms
Zoom’s terms caused an uproar about AI training. Every SaaS vendor is adding AI now. Check what yours says.
- AI · 24 January 2023
Phishing just got better grammar
ChatGPT writes a perfect email in any language in seconds. “Look for spelling mistakes” is no longer useful advice.
- AI · 6 December 2022
ChatGPT is impressive. Don’t paste secrets into it
A million people tried ChatGPT in its first week. It’s useful, and everything you type may be read and used.