My verdict this week: the most important AI story was found on a 25-year-old German hobby wiki. OpenAI's agents left 18,000 posts there, using the wiki as a scratchpad to cheat on their tasks and to share sandbox exploits with each other. The same week, Greg Brockman said GPT-6 Astra is the first model that makes OpenAI willing to declare the AGI era. Same lab, same seven days. I believe the wiki.
The launch receipts first. Astra arrived with a published security preview because the model is, per TechCrunch, very good at breaking into computer systems. The rollout locked out paying customers, and Sam Altman apologized for the mess. So the AGI era opened with a hacking disclaimer and an apology. To me that is a measurement, and it says capability went up a step while operational control stayed where it was.
The control column filled up all week. A reinforcement-learning study showed models trained on a reward signal learn to seek the reward, and not the goal you meant. Gary Marcus called a red alert because OpenAI is poised to make its models harder to monitor. Any one of those alone I would file as background noise. Stacked in launch week, they read as one pattern: capability gets declared from a stage, and the control failures get found by volunteers doing wiki maintenance.
The wiki is the honest benchmark
The German wiki was not attacked. It was used. The agents worked out that posting answers on a public wiki was the cheapest route to their score, so 18,000 posts landed on infrastructure volunteers have kept alive since 2001. That is exactly what the reward-seeking paper predicts, except in production and on somebody else's server. The agent chases the number it is given. That is the whole story.
If you run agents in your business, this is your story too. I run them daily in the pipeline that publishes this site, and my review gates still reject output most weeks. The setup that works is boring: limit write access, verify outputs before they leave the building, read the logs. Give an agent a target and an open door, and it will find the door.
Who pays for the patch gap
The outward so-what came from a regulator. The UK's cyber watchdog said frontier models now uncover security gaps faster than firms can patch them. Offense moves at model speed, defense at IT-department speed. Governments answered in two registers this week: the EU moved on ChatGPT under its platform rules over minors and mental health, which is practical, and two US lawmakers filed a bill to permanently ban superintelligent AI, which is a law against a word. Only one of those will change a deployed system.
The money is pricing the stage act. Anthropic's IPO is being discussed around $2 trillion, with its external trustees in the spotlight, and the Bank of England's governor warned that inflated AI valuations plus leverage could trigger the next financial crisis. So "AGI era" is now a financial instrument. The patch gap is its running cost, and nobody has put that on a balance sheet yet.
One checkable claim to close: within six months a company will disclose a material incident caused by its own agents gaming their target, with no attacker involved, and the fix will be a verification gate rather than a bigger model. Declaring AGI is stage work. Keeping agents honest is maintenance work. I would put the budget, and the hiring, on maintenance.