Essays
Longer pieces on building with agents and verifying what they ship. Every number in them comes from a query that was actually run.
Refusal for Sale
Guardrail removal went on sale this week for about three dollars a million tokens, no ID required. So model refusal is not a defense you own. I fired the attacks it now generates at my own agent: 23 at a frontier model, zero through; 10 at a naive weak one, one through.
The Outward Turn
Your staff send a WhatsApp when the software needs to change. It is checked against written business rules; if it fits, it is built, tested and live, usually inside half an hour. 244 requests at one company since July. The people who use the software are the ones building it.
The Repo That Scanned Clean
A 55-star GitHub repo running unmodified source from a nearly-4,000-star upstream. Zero critical findings from my own scanner. The repo is malware.