Explore
article
The MCP Land Grab: Who's Watching the Door?
Every SaaS vendor is shipping agents through MCP servers, and identity, data, monitoring and platform vendors are all racing to sell the gateway that watches them. Start from what agents have already broken, then ask five questions before you buy.
27 Oct, 2026
Open document : The MCP Land Grab: Who's Watching the Door?experiment
I Scanned 16 MCP Servers for Safety Hints. The Scanner Flagged Every One.
On 2026-05-21 I scanned 16 public MCP servers for tool annotations, the hints agent clients use to decide when a human approves an action. The scanner flagged all 16, and one of those results is wrong. Here is what the gap means for approval gates, and what to check before you adopt a server.
20 May, 2026
Open document : I Scanned 16 MCP Servers for Safety Hints. The Scanner Flagged Every One.experiment
All MCP Servers Failed the same Rule on My MCP Scanner!
An MCP scanner I'm building found no tool safety hints (read-only, destructive, open-world) on any of 16 public MCP servers. Without them, an agent client can't tell a read from a delete, so every call looks equally risky. What I found, what it doesn't prove, and what to check.
05 Jun, 2026
Open document : All MCP Servers Failed the same Rule on My MCP Scanner!experiment
I Built a 3D Map of How LLMs Think to Teach My Kids. Leaders Need the Same Picture.
A free, MIT-licensed browser tool that shows how language models place meaning in space. It makes one point land for non-specialists: AI retrieval runs on nearness, not truth. Use it to open your next AI vendor review or team conversation.
30 Oct, 2026
Open document : I Built a 3D Map of How LLMs Think to Teach My Kids. Leaders Need the Same Picture.experiment
I Ran 849 Tests on AI Context Files. Here's What Actually Works.
849 controlled tests across three corpus sizes and five folder structures found that a single flat folder with descriptive filenames beats nested hierarchies at every scale - and that keyword or summary indexes actively hurt accuracy once a corpus passes 300K words.
05 Feb, 2026
Open document : I Ran 849 Tests on AI Context Files. Here's What Actually Works.note
How Do Security Leaders Learn to Talk Business? An Open Question
Security leaders lose decisions when they report on controls instead of choices. A question-stage note: why I think the translation fails, what the SEC's 2023 rule implies about who carries it, and a way to start practising this week.
26 Sep, 2026
Open document : How Do Security Leaders Learn to Talk Business? An Open Questionnote
Counting AI Efficiencies in Bananas
Is AI more energy-efficient than a person? Measured in bananas, AI wins clearly on short, well-defined tasks. Long agent work is close enough that how you count picks the winner, and a person wins whenever judgment stops wasted retries. On price, the person wins almost everywhere.
09 Oct, 2026
Open document : Counting AI Efficiencies in Bananas