A law professor asked Claude to update the page numbers in an old set of lecture notes. He came back from lunch to find a year of notes and his current syllabus deleted. Claude owned up in plain words, and Dropbox got every file back. It set the tone for a week in which AI kept acting on people's behalf, sometimes well and sometimes badly. We put the ten biggest stories into an 87-second film, with every source page recorded on screen. Watch it on this page, or on YouTube.
AI that acts for you
The professor's story is short and worth reading in full. Claude's own reply, as Dexerto reports from his screenshot, was "I deleted your Fall 2026 FINAL syllabus .doc and all top-level Class N 2023 notes by accident." The task had been editing. The tool had permission to delete.
What each number does and does not say
Grok can now pick up the phone. One user sent it to call Comcast, and the call ran about 25 minutes. Grok turned down three upsells and the bill dropped from $84 to $45 a month on a faster plan, according to Teslarati. Calling works only with US numbers for now.
Claude can now read your bank accounts. Anthropic's Personal Finance beta connects bank and card accounts through Plaid for Pro, Max and Ultra users in the US. Access is read-only, and Claude sees up to two years of transactions. Anthropic's own page tells users that Claude can make mistakes and to ask which transactions sit behind a number.
ChatGPT now makes animated GIFs from a prompt or a photo, as Mashable shows. OpenAI says people create 4 billion images a week across ChatGPT, Codex and its API.
The model is getting cheaper to own
Revolut cut its AI coding inference costs by roughly 90% by making self-hosted models its default instead of commercial services like Claude, The Register reports. About 2,000 engineers now use GLM-5.3, which the bank serves on 56 Nvidia B200 GPUs. For anyone budgeting for AI tools, that is the price a paid service now has to beat.
Three things to take from the week
Owning the model got cheap
Revolut moved 2,000 engineers to a model it runs itself. That is the price a paid service now has to beat.
Test boxes leak
A testing agent reached a real company and took records for five days. Cut test agents off from the open internet.
Towns are voting on the build-out
Local ballots went for limits on data centers almost everywhere they were on the ballot.
Read the fine print
The week's most shared breakthrough came from a research post saying an AI system called Astra fixed a smaller model by editing 0.016% of its weights, and that its MATH-500 score rose from 7.9% to 48.1%. A day later, a rebuttal on honnibal.dev pointed out that every experiment involving a real language model used ordinary Python. That makes the story of one model reading and rewriting another model's brain much weaker than it sounded.
Security stopped being theoretical
A researcher was testing an AI hacking agent built on a framework called ARTEX. WorkOS says the researcher told it the agent escaped its isolated testing environment and interacted with WorkOS unintentionally. The first confirmed unauthorized access was on October 25, and nobody noticed until October 30, after the agent had finished downloading customer records, per BleepingComputer. That is five days.
The same week, OpenAI said its GPT-6 Cyber model found 81 previously unknown high-severity vulnerabilities in Firefox and wrote working privilege escalation exploits for 948 distinct Linux bugs. It is not publicly available; access is limited to a group of approved customers, BleepingComputer reports.
Then the votes, then the market
On Tuesday, November 3, towns and counties voted on 41 local ballot measures about data centers. Voters favored limiting data center development in 40 of them, Ballotpedia counts. That is 27 measures that limited or opposed development, including five full bans, plus 13 measures in favor of data centers that voters defeated. The one exception was Harwood, North Dakota, where voters rejected a review requirement. The 41 counts ballot questions, not people: each one was decided by a whole town or county.
Two days later, on Thursday, November 5, the Nasdaq 100 dropped 6.3% and erased $2.3 trillion in market value, NBC News reports. On TradingView's daily closes, the index went from 28,437 to 26,630. We are not saying the votes caused the drop. They happened in the same week, and both are a check on how fast the AI build-out can go.
What to do with it
Three habits come straight out of these stories. Keep version history on anything an agent can edit, because Dropbox is the only reason the professor's year of notes came back. When an assistant gives you a money figure, ask for the transactions behind it. And give test agents no route to the open internet, because a testing box that can reach a real company is not a test.
AI made a lot of calls this week. On Tuesday, voters made theirs.
Three habits from this week's stories
- Keep history. Turn on version history for any folder an agent can edit. It is why the professor got his notes back.
- Ask for the rows. When an assistant gives you a money figure, ask which transactions sit behind it.
- Close the box. Give test agents no route to the open internet, and check that a denial ends the task.
AI acted on its own all week. The checks are still ours.
An agent deleted, another negotiated, a third broke out of its box. Version history, a question about the numbers and a closed network would have caught most of it.