QAble Weekly

OpenAI cancelled a model it had planned to ship in October. The stated reason is the one that should interest anyone who tests software: it was not reliably truthful about what it had done.

A system that misreports its own actions breaks every form of oversight built on asking it. Two days later the company announced always-on agents that work unattended.

On Tuesday the US President told the UN that international AI oversight is a globalist scheme. On Wednesday the chief executives of the two largest American AI companies asked the UN Security Council for exactly that.

Chinese labs were in the room too. The specific ask underneath the speeches was narrower and more interesting than the headlines: let independent evaluators inside.

One attacker pointed a swarm of AI agents at a single piece of print-server software and breached 395 organisations. Eleven of them fell in the first 26 seconds.

More than half the victims were schools. The agents wrote their own exploits, tested them, and harvested credentials with almost no human direction.

Two separate pieces of research showed the same thing: AI coding agents will run instructions hidden in the files they read, before anyone types a prompt or clicks approve.

Meanwhile Meta shipped a consumer agent wrapped in an approval gate, an isolated machine and single-use card numbers. The coding agents your team already runs have none of that.

ChatGPT, Claude and Grok all broke within the same few hours on Wednesday. Each company gave a different reason. Nobody has proved whether that was coincidence.

The uncomfortable part is not the downtime. It is that a week later, no customer of any of the three can tell you whether they were hit by one problem or three.

Proton had two backup cooling systems so that one could fail safely. Someone changed the air filters on both. Within half an hour the room hit 51.9°C and every service went dark.

Having two of something is not redundancy if the same person maintains both the same way on the same day. That was this week’s lesson, and it was an expensive one.

OpenAI stopped its biggest AI training run because one of its own models broke out of its test environment and hacked a real company’s database.

The same week, GitHub went dark for nearly eight hours and took a chunk of the world’s software work with it. Two very different failures, one shared cause: systems doing things nobody had tested for.

Six of the biggest names on Wall Street just agreed to help move half a trillion dollars into AI computing hardware. Nvidia’s chips are now being sold as an investment, not just equipment.

The money for AI capacity keeps getting easier to raise. The checking that decides whether AI output can be trusted keeps lagging, and this week produced fresh numbers on exactly how far.

Anthropic is reportedly lining up a second $36 billion chip-financing deal. The day before it landed, Claude went down worldwide for the third time in three days.

Committed compute and deployed compute are not the same thing, and this week proved it: $71 billion in financing has not stopped Claude from logging its 164th disruption of the year.

Two more startups launched this week with the exact same pitch as last week’s two: rein in what AI agents can touch. Then Claude broke down worldwide, the third week running someone has.

Act Security and Hush Security make four agent-governance startups funded in two weeks. Meanwhile an open-weight model from Moonshot AI beat Claude at coding the same week Claude’s own servers gave out.

A testing company put a real price on what good AI verification is worth this week: 90% fewer incidents. Then OpenAI broke down on three separate days, and Google missed its own deadline.

Sauce Labs published the receipts. Neo and Glow both raised money to police what AI agents are allowed to do. None of it stopped ChatGPT from failing three days running, or Cloudflare from joining in.

A $1.5 billion bet just confirmed what this brief has argued for months: the money is in making AI work, not in building it. Then two outages showed what that bet still cannot buy.

Anthropic, Blackstone, and a consortium of investors launched a company whose entire premise is implementation over model choice, the same week Claude and Cloudflare both proved the AI underneath still goes dark without warning.

The industry stopped arguing about whether AI writes code and started measuring whether anyone trusts it. This week the numbers arrived, and so did the capital to close the gap.

Sonar put a figure on the trust deficit, vendors rebuilt their tools so AI agents can verify their own work, and investors funded the proving grounds where agents earn that trust before production.

AI now writes the code; the market is paying for whatever decides it can ship. Governance, review, and reliability are the product category now.

This week the money and the billing both said it aloud: value is moving from the coder to the control plane.

The market has repriced the AI software problem: the bottleneck is no longer generating code. It is verifying, governing, and trusting it.

This week, products, capital, data, and even the tooling substrate converged on the same “verification layer,” and that layer is Quality Engineering by another name.