Earlier Essays
The essays that have rolled off the front of the Writing index. The most recent work lives there; everything earlier is collected here, newest first.
-
The Human Approved It. Which "It"?
A human approved the deployment at 2:00; the agent executed it at 4:17. In between, the artifact was rebuilt and the target set changed — and the approval record still said APPROVED. Human-in-the-loop stops meaning anything in that gap. The control that matters is the binding that makes an approval apply to this action and no materially different one.
Read the essay → -
The Control Wasn't Defeated. It Wasn't There.
Anthropic's August 2026 Risk Report documents safeguards that were absent, misconfigured, or simply not on the path that ran. The recurring pattern is not a defeated control but assurance-state divergence: the running system drifting from the control model that governs it. A defeated control leaves a mechanism to investigate; an absent one leaves the belief where the evidence should have been required.
Read the essay → -
The Assumed Control
Some failures that appear to show a control being defeated are actually the discovery that the control everyone relied on was never present. The belief occupies the place where the work should have happened, and the absence conceals itself, because everyone can point to where the protection is and the place they point to is a description rather than a mechanism.
Read the essay → -
The Wall Was Never There
When an autonomous system does something it was not supposed to do, we call it an escape. Usually nothing was broken through, because nothing was built to break through. Autonomy relocates execution; it does not relocate accountability, and narrating the event as a story about the machine's power is how that attribution goes missing.
Read the essay → -
When the Machine Drives the Nails
Execution got cheap this year. The interesting question is not what that gives us but what it quietly takes away, and from whom, because the capability the model absorbed is the same one that used to build the judgment now required to direct and repair it.
Read the essay → -
You Cannot Classify Your Way Out of an Algorithm
Google's quantum ECDLP disclosure reduced harm but did not contain the result: rediscovered in nine weeks, improved on an open scoreboard, and its zero-knowledge proof forged through a software bug. Why the disclosure, not the nine-minute headline, is the architect's lesson.
Read the essay → -
Your Browser Is an Environment, Not a Tool
Three recent browser changes, FROST, Gemini Nano, and Manifest V3, are each defended in the language of your security while transferring control to the vendor. The justification is the tell.
Read the essay →