What moved in tech this week — and what it means for EdgePointe

A thing we verified last week was reversed eight days later — and that is the business Bet intact

Our bet has never been on which AI company wins. It is that somebody has to stand between a small business and software that changes underneath it every few days. This week made that case better than any pitch deck we could write.

In five days, the tool at the centre of our own build shipped eight releases. Five of them fixed security holes in how permissions are enforced. Two of them fixed leaks that were printing passwords into log files. And one of them reversed a fix we reported to you last Monday as closed.

None of that is a scandal — it is a fast-moving product behaving like one. The point is what it costs a business owner who has nobody watching. A verification is good for about a week. That sentence is either a nuisance or a service line, depending on who you are.

Finished work is still parked — eleven days now — and one date lands Friday Clock running

Nothing on that list is hard. Everything on it is old.

A five-day repair wave — and one part of it we cannot just upgrade our way out of Act now

Between Sept 15 and Sept 18 our core tool fixed five separate flaws in how it decides what the AI is allowed to do, plus two credential leaks. Most of that we collect for free by updating.

• The serious one: a destructive delete command could be hidden inside a wrapper and slip past the safety check. That is exactly the shape a malicious instruction would take.
• The one that needs a hand: two separate bugs were printing passwords and tokens into error messages and log files until Sept 17. Updating stops future leakage — it does not un-write what is already in the logs. Anything that sat in those places gets rotated.
• Free protection we did not ask for: two changes landed on Sept 18 that harden exactly the risk in our website-cloning work — untrusted text pulled off a stranger's website can no longer masquerade as an instruction from us.

Vantage is fine — and Cloudflare just handed us a twenty-minute security upgrade All calm

Nothing on our platform broke, is breaking, or is scheduled to break. Three notes, in descending order of usefulness.

A paid assessment we can sell without building anything first Act now

Cloudflare made a feature generally available this week that hands us a service we have been describing for a month: showing a client what AI tools their staff are already using without permission.

The detail that makes it sellable: it produces findings without anyone having to configure a policy first. That is the exact shape of a paid assessment — we turn it on, we read the report, we walk the owner through what we found, and the remediation that follows is the recurring work.

Why now, and not next quarter: most of our clients have not deployed an AI agent yet, so governing agents is not the near-term product — visibility into the AI already in the building is. Every one of them has staff pasting company information into a chatbot today. Nobody has told them.

A second, smaller Cloudflare change makes client-owned accounts easy to create — which is a better answer to "who actually owns this website" than anything we have said to date.

Our phone AI's worst failure just got fixed — by the same vendor now selling one of its own Fix + watch

The single worst thing our receptionist could do is fail to pick up when a venue gets busy — a bridal show weekend, an ad burst, a post that takes off. Until this week, hitting that ceiling meant a lead simply never got through. That is the outcome that ends a contract.

• The vendor shipped queueing with branded hold audio and a bounded wait. Callers hear the venue's own recording instead of nothing.
• Hold time is not billed.
• It is switched off by default — so we do not have it today. Turning it on and uploading each venue's own clip is small work that reads as bespoke to an owner.

The uncomfortable half: the same vendor's own AI receptionist product now sits as a top-level section in their documentation, beside their core platform. They are not experimenting. Our defensible line was never the voice — it is the venue-specific booking logic, where the lead goes next, and a human being answering when the AI cannot.

A 541,000-judgment study says the test we designed last week would have lied to us Redirect

We have been building a way to score whether a website we rebuild is actually better than the one it replaced. Last week's briefing told you how to run the control. A study found this week says that control does not work.

Scale first, because it matters: 21 different AI judges, nine vendors, about 541,000 individual judgments. Three findings that cost us something:

• An AI judge can be perfectly consistent and still be reading the wrong thing. Our plan was to check it against itself. Self-agreement does not detect the error — the control has to be a question where we already know the right answer.
• Percentage agreement overstates these scores by roughly a third of the scale. Every quality number we have quoted internally is inflated.
• Verbosity is not the problem we assumed — which cancels a build. That is the only free thing on this card.

Separately, a second study on voice agents says judge the phone AI partly by machine and partly by human — and the two things it says a machine cannot yet score reliably are did the agent recover when it got confused and did it miss something it should have caught. Those are the two a venue owner cares about most.

Two decisions need Bob and Niven this week — and one has Friday attached Decisions

Two things Paul does without you this week: rotate the credentials and narrow the deploy key (one sitting), and turn on hold queueing for the venues.