One disclosure in 431 was ever used
A measured ratio behind the patch backlog, a spend report that finally names whose work it was, and a service tier taken out of the hands of the people who read it.
How we Organise
The same token count covered a code review and a one-line summary
Cloudflare has added task context to the usage reporting in AI Gateway. The gateway now sorts conversations by the kind of work behind them, coding, research, summarisation and the rest, and names the user, the agent and the application driving each one. It also marks the cases where the model chosen looks more capable than the job needed. Two further views ride on the same signals, one listing requests a cheaper model could have served and one routing them there automatically. Cloudflare publishes no sample and no figure for how often the mismatch occurs.
We tell a finance team that nobody cuts a bill they cannot attribute, and an AI bill has mostly arrived as one number from a vendor.
Costs that nobody owns cannot be reduced.
What changed this week is that the attribution ships free instead of arriving as a consulting engagement. Whether the spend then falls is a separate question. Cloudflare offers no before-and-after and says plainly that the view recommends no replacement model, so the saving is left with the customer to find.
How we Build
The service tier is now read-only in the place people read it
GitHub has put a repository’s business context under outside control. Ownership, service tier, lifecycle stage and compliance status can be pushed in from a configuration database or an internal developer portal. An API of their own keeps them current. GitHub refuses edits to them in its own interface, so two systems cannot both claim the same field. Governance rules can fire on the synced values. Port.io is the first partner and the feature is in public preview.
Our position is that a classification rots, and a tier nobody re-dates stops describing anything.
A classification is perishable, and a label with no review date is a decision the firm has stopped making.
GitHub has met the same problem and answered it with continuous synchronisation rather than a review date. That is the better answer wherever the owning system is alive. Where it is not, it is no answer at all. A tier synced out of a database nobody maintains is exactly as stale as one typed in by hand, and being read-only now makes it look authoritative.
How we Assure
Patch everything and you work through 431 to reach one
Google’s threat intelligence group counted what the disclosure flood actually contains. Monthly disclosures doubled inside 2026, from just over five thousand in January to nearly eleven thousand in August, and exploitation rose alongside them. The proportion ever observed being exploited is 0.23 per cent, which is roughly one disclosure in 431. Descriptions merely containing the words Linux Kernel produced about five thousand of those identifiers, with no exploited zero-day among them. The report’s own recommendation is to give up unprioritised mass-patching for triage driven by threat intelligence.
We judge a control by what it costs set against what it answers, and patching everything is the clearest case of one that fails.
A control heavier than its risk produces use nobody can see.
The argument for triage used to rest on how much a team could get through in a month, which is a capacity argument and loses to a determined auditor. It now rests on a measured ratio. What the figure cannot say is which of the 431 it was. The report is explicit that at these volumes the monthly counts move on vendor release cycles rather than on anything structural.