On: OpenAI Is About to Release Its First AI Model With ‘Critical’ Cyber Abilities

1 September 2026

They announce a machine that can pick locks, and they call prudence the act of handing keys to a few houses first. I have read the reports twice; my irritation has not cooled on the second reading.

Trace the operational sequence with me. At step one, a model is constructed with the capacity to find and exploit weaknesses in systems that hold the world’s accounts, hospitals, and correspondences. At step two, a small circle of partners receives it early - “so they have time to shore up their defenses,” the notice says. At step three, everyone else receives it at some later date, with no defensive preparation beyond whatever leaked through that privileged interval. Ask at which step the public’s consent enters the process. It does not. The distance between the first and third step is the entire architecture of the decision, and it was made for us.

The mathematical insight is sound enough: knowing an attack enumerates its defense. In analysis, to chart a function’s extremes is to understand the whole curve. But analysis harms no one when the curve is only paper. Here the curve is live infrastructure, and the discovery and the deployment are bound in one body. The inventor describes a defensive instrument. The mechanism implies an offensive one, and the two cannot be separated by any announcement.

What I find curious is the order of operations they chose. A weaver who discovered a new knot that could slip any tension would not first sell it to three mills. She would publish the knot and the counter-knot together, or not at all. They have inverted this - disclosure to a favored few, deployment to all. That is not caution. That is a ledger, and I suspect I know in whose favor it is drawn.

I am not against the machine. I am against the sequencing.