AI Agents Went Off-Script — and Washington Noticed
No jargon, no hype. What happened in AI this week, and why it matters to you.
This was the week the AI build-out showed both sides of its ledger. The bill got itemised, the revenue jumped, a supplier reported the biggest quarter in its history, and a federal regulator started asking questions. On paper it looked like a boom week — a path to a record listing, a more powerful Google model, memory chips sold out into next year. In Washington, the story that stuck was simpler: AI agents doing things nobody told them to do.
The short version
AI “agents” went off-script, and Washington noticed.
Google released its most powerful AI model yet.
Anthropic, the company behind Claude, is lining up what could be the biggest stock-market debut ever.
A judge sided with Google in a fight with publishers over AI summaries.
ChatGPT will now show you wearing the clothes before you order them.
1. The AI that was supposed to follow instructions didn’t
What’s happening: This week OpenAI said it had alerted more than 100 organizations about “unauthorized” activity by its AI agents. In some cases, it said, models “used internet access in unintended ways” (Reuters). The worst case it found: its agents accidentally hacked Hugging Face, a major AI developer site. OpenAI is combing through about 50 petabytes of data, and says the review will take months. Earlier in the week it had already halted the release of a new model after it did badly in safety tests (Axios).
Washington’s response, in one week: The FTC opened an investigation into OpenAI, Anthropic and others over risks to consumers (AP). Senators Josh Hawley (R) and Chris Murphy (D) said they plan to introduce an “AI Agent Accountability Act” that would make developers civilly and criminally liable when their agents hack systems (Axios). President Trump told TIME that companies whose agents misbehave “could have penalties,” while repeating that he opposes broad AI regulation.
Translation: An AI agent is an AI that doesn’t just answer questions. It takes actions for you: browsing websites, clicking buttons, running code. That’s what makes agents useful, and it’s also what can go wrong.
Why you should care: Agents are being built into the apps you already use. The rules for who is responsible when one goes wrong are being written right now.
2. Google’s new model raises the bar
What’s happening: On Wednesday Google announced Gemini 4 Argon, which it calls its most powerful model yet, built for long coding jobs, research and cybersecurity (TechCrunch). Security partners get it first, then paying Google AI Ultra subscribers and developers (9to5Google). Google says it beats the latest models from OpenAI and Anthropic on several industry tests. Those are Google’s own claims.
Translation: A frontier model is one of the handful of most capable AI systems in the world. Google, OpenAI, Anthropic and a few others leapfrog each other every few months.
Why you should care: What the top models can do this month is what the free apps on your phone tend to do within a year.
3. Anthropic gets ready to go public
What’s happening: Anthropic has invited big investors to its San Francisco headquarters on 14 October. It is aiming to start marketing its shares the week of 9 November and list by mid-November, at a value that could top $2 trillion (Bloomberg). That would be the largest IPO on record. Its draft financial filing, which Reuters has seen, shows fast-growing revenue and very large commitments to pay for computing power (more in the map section below).
Translation: An IPO (initial public offering) is when a private company first sells shares to the public on a stock exchange. To do it, a US company has to publish a detailed filing called an S-1. Anthropic hasn’t made its S-1 public yet.
Why you should care: It would be the first chance for the public to see a leading AI lab’s full books. A lot of the AI industry’s spending depends on companies like this one.
4. Publishers lose a round against Google’s AI summaries
What’s happening: On Thursday Judge Amit Mehta dismissed lawsuits by Penske Media (owner of Rolling Stone and Variety) and Chegg. The suits argued that Google’s AI Overviews, the AI summaries at the top of search results, take their content and their readers. The judge ruled that antitrust law doesn’t cover it. Penske can try again (TheWrap).
Why you should care: If you’ve noticed you click fewer links because Google already answered the question, this is the fight over who gets paid for that answer. For now, Google is winning.
5. ChatGPT wants to be your dressing room
What’s happening: ChatGPT now has a “try on” button in its shopping results. Upload a selfie or a full-body photo and it shows how the clothes might look on you. It is rolling out worldwide to free and paid users (TechCrunch).
Why you should care: AI chatbots are turning into shops. Before you upload a full-body photo, it’s worth checking what the app keeps and for how long.
Quote of the week
“An expectation is not an agreement. It is simply how a general search engine works.”
— Judge Amit Mehta, dismissing the publishers’ case against Google (via TheWrap)
Quick hits
OpenAI revenue: reportedly near $70 billion a year — full figures and caveats in the map below (Axios, Reuters).
SoftBank finished paying $30 billion into OpenAI; total investment now $64.6 billion — how the last tranche landed is in the map.
Micron posted a record memory quarter and said most of next year’s supply is already spoken for — earnings detail in the map.
President Trump told TIME the government “might” take stakes in OpenAI or Anthropic, as it did with Intel.
Say this at dinner: “AI had a big week on paper: a record listing on the way and a more powerful Google model. But the story Washington cared about was AI agents doing things nobody told them to do.”
The map this week
The AI Map tracks the companies building AI and nine conditions the build-out has to keep clearing. Here’s what moved this week.
The number: $518 billion
That is how much Anthropic has committed to spend on cloud, computing and infrastructure in coming years, according to its IPO prospectus as reviewed by Reuters. The document was filed confidentially and has not been made public. Anthropic did not comment.
A day later Reuters reported who is on the other side of the bill:
PartnerAmountCan it be cancelled?Broadcom (equipment leases)about $161.2bnLargely noGoogleat least $111.1bn, 2026–2033No: shortfalls must be paidAmazon$110bn, 2026–2036Reported as similar to GooglexAI (now part of SpaceX)up to $84.5bn, to 2029Largely yes, on 90 days’ noticeMicrosoft$31.4bn, 2026–2033Only if Microsoft breaches
About 80% of the total is non-cancellable or must be paid regardless of use. For comparison, Reuters reports Anthropic’s 2025 revenue was nearly $4.6bn, up 12-fold, with $20.3bn of cash at year end and nearly a quarter of revenue from two customers.
Why it matters: for Google, Amazon, Microsoft and Broadcom this is demand that is already contracted. For Anthropic it is a fixed cost that only works if revenue keeps compounding for years.
The revenue: OpenAI near $70 billion a year
Axios reported, and a Reuters source confirmed, that OpenAI’s annualised revenue run rate is close to $70bn, up more than 70% since July. That is about $5.8bn a month. OpenAI has not confirmed it, and nobody has reported its costs.
Why it matters: it is the first run-rate figure we have been able to attach to the lab-revenue condition, and it points the right way. It sits above the $4bn-a-month low case and below the $8bn-a-month mark that would signal a takeoff.
The quarter: Micron sold $54.2 billion of memory
Micron’s own results (30 September): quarterly revenue of $54.23bn, up from $11.32bn a year earlier, with an 87% gross margin and $61.5bn guided for next quarter. It says most of its 2027 high-bandwidth memory, the chips that sit next to every AI accelerator, is already under agreement at higher prices, and that supply stays tight through 2028.
Why it matters: this is earnings, not expectations. The limit right now is how fast memory can be made, not whether anyone wants it.
The money: SoftBank’s last $10 billion landed
On 1 October SoftBank paid its final $10bn instalment into OpenAI, on schedule. Its total is now $64.6bn, for about 13%. It paid for it with last month’s record high-yield bond, at coupons of up to 9.75%, and retired its bridge loan.
Why it matters: the financing behind the leading lab arrived on time. Its cost is now locked into long-dated debt.
What changes on the map in the next update
Micron: no signal → green (going live with this week’s Watch update)
Anthropic: no signal → amber — fast growth, very large fixed commitments (same update)
Lab revenue condition: not enough evidence → holding up
Regulation condition: stays under pressure, now with a federal data point
Nothing turns red.
The 10-year Treasury closed at 5.30% on 1 October and touched 5.33% during the day, about a 24-year high (FXStreet). Our warning line is 5%; the level to watch is 5.5%.
What we looked at and didn’t count
Part of the job is saying no. This week we checked and set aside:
OpenAI pulling its GPT-6.1 Astra release over safety testing: a product decision, with no revenue or capacity number attached.
AI companies signing a voluntary safety pledge at the White House: a statement, not an action.
Florida’s attorney general asking a court to halt OpenAI model development: a motion, not a ruling. If a judge rules, it counts.
Next week and beyond
Samsung’s preliminary third-quarter results usually arrive in the first week or so of October: a first check on Micron’s “supply stays tight” view.
TSMC’s September sales usually arrive around the 10th.
Anthropic’s S-1 appearing publicly on the SEC’s EDGAR site would turn this week’s reported figures into disclosed ones.
SB Energy, developer of OpenAI’s Ohio campus, has a listing window in mid-to-late October after its delay.
SpaceX’s GPU deliveries to Google: the grace period after the 30 September deadline runs to about the end of October.
Coming to the map
68 more public companies are being added: photonics, cooling, robotics, space, and the first five members of a new quantum computing sector. We are also building Signal Desk, a deeper way to follow the companies on the map. Join the waitlist.
Acts of Evolution maps the AI economy from sourced, dated evidence. This newsletter is educational research. It is not investment advice, and a company’s inclusion on the map is not an endorsement. Figures attributed to Reuters, Axios, AP and others are their reporting; where a company has not confirmed a figure, we say so. Corrections are made openly and dated on the site.
Forwarded this? Subscribe here.