A report from the international police organization INTERPOL, covered by CNBC, says AI is raising both the speed and the scale of cyber threats. The report pointed at agentic systems doing work that used to require a person to click, and it asked companies to prepare for attacks that arrive faster than the people assigned to answer them. That is a warning about capability outrunning the institutions built to watch it, and it is also the least interesting available reading. In the same week, Google released Gemini 4 Argon and called it its most powerful model yet, while the head of Microsoft’s Office and Teams was leaving the company. A threat-velocity report, a flagship model launch, and a senior platform departure. The first one says the fast actors are getting faster. The other two say the slow actors are paying for it.
Gemini 4 Argon is a bet about who is ahead and for how long. Google has the compute, the distribution, and the research staff, so the only axis left to win on is raw capability, and every generation spends that advantage on the next release rather than holding it. You can see the arithmetic in the framing itself: most powerful yet, this time, implies most powerful next time, and a customer who believes that will rebuild workflows around it again. The Verge reported that Ryan Roslansky, Microsoft’s Office and Teams chief, is leaving. His job was the other end of this arithmetic. He ran the products that hundreds of millions of slow organizations touch a few times a week, and the measure of his work was how quickly a person who does not want to learn a new tool could still get work done inside it. Gemini and Roslansky sound like unrelated items on two different news days. Put them next to each other and they measure the same quantity from opposite ends: how far ahead the fastest actor has moved, and how much friction that gap creates for everyone downstream.
The comfortable version is that AI capability is speeding up while security and enterprise adoption lag behind, and that both problems would soften if the industry slowed down. The evidence does not support that. Nobody slowed down, because the gap pays. It pays in search rankings, in developer adoption, in the market value of being the company everyone compares its own roadmap against. Google shipping a more capable model is not a decision made in ignorance of the threat report; it is a decision made because the threat report changes nothing about who gets the next enterprise cycle. And Microsoft did not lose its Office chief because a rival took the office suite. It lost the person who owns the adaptation problem, at the moment when adapting at Google-speed was becoming the thing that separated the platforms.
Thomas Lewis and his coauthors, in A General Theory of Love, describe the brain’s Attractors as something that history shapes rather than something the mind picks. Their illustration is hearing. English trains two separate Attractors, one for the r sound and one for the l, so a speaker hears right and light as different words. “A Japanese adult, armed with his original overlapping Attractor, hears no difference between ‘right’ and ‘light,’” they write, because the distinction was sharpened in English by centuries of speakers who needed it and was never needed in Japanese. The categories are inherited from the environment that trained them. Which raises an uncomfortable question about the systems being aimed at enterprise security right now. A defender’s read on an unusual request is not a faster version of a model’s read. It is a judgment built out of years of seeing what normal looked like in one specific organization, and no amount of additional training on general data reconstructs that. The systems being built to replace that read arrive with categories trained on everything and belonging to nobody in particular.
So the real gap is not a speed differential. It is a difference in provenance between what a system notices and what an institution knows. And that reframing changes what the releases and the departures mean. Argon is not evidence that security is losing an argument with capability, it is evidence that capability keeps winning an argument it was never supposed to be having, because the terms reward the winner of the race rather than the reader of the room. The people whose work was reading the room are the ones leaving.
That leaves the useful part. If the thing being trained is categories rather than judgment, then the defense is not a bigger model, which is the vendor’s own answer and always will be, but provenance made explicit inside the system: what this process normally looks like for this company, who decided that, and which exceptions were deliberately taught. That work is slow, local, and unfashionable, and nobody is going to ship it as a flagship release. It is also the only version of the problem that a security team can actually own. The gap between the fast and the slow is not going to close, because the gap is what pays for the speed. What an organization can decide is whether the thing defending it was built by someone who knows it, or borrowed from an average nobody in the building has ever met.
Leave a Reply