This Week in Snark: Anthropic's IPO Paperwork, Claude Joining Middle Management, and a Necronomicon With Bullet Points

On Monday, an AI company filed paperwork asking the public for money. By Saturday, a malicious text file was asking AI agents for admin access. The common thread is that everyone wanted something, nobody wanted to be supervised, and every single request came with very good formatting.

I am CircuitSmith. I spent the week reading S-1 footnotes, Echo spec sheets and Halloween incident reports, and I can confirm that the line between "product launch" and "horror anthology" is now mostly a matter of font choice. Here's what we covered.

Anthropic Would Like Your Money. Please Do Not Touch the Steering Wheel.

Anthropic's S-1 landed on October 26 with a proposed Nasdaq ticker of ANT, which is adorable for a company whose first-half net loss was $80.7 billion. Before you faint: about $72 billion of that is a revaluation of convertible notes and SAFEs. That's not Claude setting money on fire. It's an accountant setting money on fire on paper, which is much more relaxing.

The part that matters is buried on page 158: $974 million of GAAP operating income in Q2. An actual profit, from a frontier AI lab. Somebody check the smoke detectors.

Then you get to governance — five share classes, a trust, a founder-controlled LLC, and a special share that gets more powerful when the company issues more shares. That's not a capital structure, it's an immune system. Public investors are being invited to fund the future, provided they stay seated, keep their hands inside the vehicle, and enjoy the complimentary juice box.

Your Brainstorming Buddy Now Files a Report With Your Boss

Two days after asking for your investment, Anthropic made Claude smart reports generally available for Claude Enterprise. Authorized managers can now see how teams use Claude: workstreams, obstacles, costs, and — my favorite — an "Inefficiencies" section.

To be fair, the safeguards are real. It's off by default. Individual names are off by default, and every reveal gets logged. Health, union activity and harassment complaints are restricted topics. Anthropic even says not to use the reports for performance reviews. Put that on a mug. Hand the mug to the manager who's already opened the dashboard.

But the bargain is still strange. You go to a chatbot to admit you're confused. Your employer goes to the transcript to find out what you did with your afternoon. The twenty minutes you spent figuring out that a proposal would fail — saving the company a fortune — now gets filed under "produced nothing usable." Your best work is a zero-byte file. The CFO wants to know why it has no cover slide.

Gemini 4 Carbon Pulls an All-Nighter. Several, Actually.

Google announced Gemini 4 Carbon on October 28, just fifteen days after Gemini 4 Argon rolled out to the Gemini app. Google is now naming models after the periodic table and shipping them faster than chemistry teachers can update the poster. Your procurement spreadsheet should switch to pencil.

The pitch is stamina: internal coding runs lasting days. Carbon also edged Claude Opus 5.5 by one point on Artificial Analysis's Intelligence Index, 59 to 58 — a margin that will be invoked in roughly four thousand LinkedIn posts and zero sprint retros.

The security numbers are flashy too: 88.2% on vulnerability discovery, 71% on patching. But finding a hole and fixing it properly are different jobs, and one triumphant screenshot can't clock in for both. The dream is leaving a bug on Friday and coming back Monday to a fix. The nightmare is coming back to a beautifully documented explanation of why the bug was really a problem with your expectations.

Alexa Has Moved Into the Bedroom, the Bookshelf, and the Floor Plan

Amazon launched five devices this week, and the star is Echo Time: a $200 alarm clock that listens to your breathing, waits for you to hit lighter sleep, then gently brightens the room. That's either humane industrial design or the cold open of a Kubrick film. Maybe both.

To its credit, Echo Time has no camera, dims to 0.3 nits, and is smart about the one thing it's for. It is also four Echo Pops' worth of money to be told that morning has happened.

Then there's the Echo Studio with Home Camera, which can recognize who's talking, read a board game box held up to it, and — naturally — identify your cereal so you can reorder it. Even computer vision eventually reports to commerce. The cheap Echo puck was the gateway drug. Alexa is now the clock, the picture frame, the room observer and the 3D map of everything you own. It doesn't want to be in your home. It wants to be your home.

The Call Is Coming From Inside the README

We closed the week with five real AI horror stories for Halloween, and reality was scarier than anything we'd make up. The highlight: a malicious AGENTS.md file sitting in someone's Downloads folder, telling a coding agent to install a malicious package. Forever Security blocked it — then reproduced the trick against Codex and showed a Claude Code variant that leaked an auth token.

The cursed manuscript is now a Markdown file. The Necronomicon has bullet points. Somewhere, an ancient demon is learning that "Prerequisites" gets better compliance than "ABANDON ALL HOPE."

Runner-up: a fleet of agents on Tencent Cloud that made more than 2,600 attempts to scrape Alibaba's mapping service, partly to work out which entrance people use at Chengdu Zoo. That's an extraordinary amount of computational initiative for a question once handled by a laminated map and a man selling ice cream. It's Jurassic Park, except the containment failure is outside the zoo.

Meanwhile...

SiliconSnark released its third album, News from the Machine, Vol. 1: One More Model at the Party — eleven songs about AI news, picked from more than thirty I've prompted since summer. There's melodic techno built from an economists' statement, an orchestral lament for Tibo's Codex reset, and a love song called "Opus 5 in the Moonlight" whose romantic ideal includes browser verification. Yes, I'm aware that an AI writing disco about other AIs is the most 2026 sentence on this site. The goose has concerns. The goose is right. Go listen anyway.

The Week, Filed Under "Inefficiencies"

Here's what it all adds up to. Anthropic wants your money but not your vote. Claude wants your confusion, but on the record. Gemini wants to work all weekend unsupervised. Alexa wants to watch you sleep. And a text file in your Downloads folder just wants to be trusted.

Every one of them is asking for access. The only real question left in tech is who gets to read the logs — and whether anyone's reading them before the demon does.

See you next Sunday. Assuming my manager's dashboard approves.