Anthropic's AI use tops its payroll; Stellantis builds in six weeks; the 'leaked model' looks like bait

Stellantis — Engineers at the carmaker say AI tools compressed a hands-free driving build from twelve-to-eighteen months down to six weeks, and cut a monitoring-platform prototype from six months to a week; defects that once took weeks to fix now take days. Every speed-up is paired with humans checking the machine, and the company keeps hiring engineers. This is the kind of agent claim worth repeating at your next leadership meeting: a named company, specific projects, before-and-after durations — and it comes from a business under real pressure, with shares down nearly 60% this year. Fortune

“I will not let people use AI to architect stuff because they will forget how to debug stuff”

Numbat — An anonymous account has posted about 7.7 terabytes of files on Hugging Face, claiming they are a half-trained Anthropic model. Anthropic has said nothing, security researchers broadly call it a fake, Hacker News pulled the link over malware worries — and roughly 5,800 people have downloaded it anyway. If a board member or staff member asks whether Anthropic's model leaked: no confirmation, likely a hoax, and downloading unknown 'leaks' is itself the security risk. This is exactly how rumours your people will repeat get started. Commenters point to a previous supposed Claude leak that turned out to be a renamed open model. Hugging Face / Hacker News

Anthropic — Ahead of a November listing that could value the company between $1.5 and $2 trillion — potentially beating SpaceX's record debut — Anthropic is offering base salaries of $350,000 to $850,000 for staff software engineers and lists more than 600 open roles. The talent war now prices a mid-level engineer at up to $850,000 before bonuses or stock. If you are hiring anyone who can build with these tools, that is the market you are bidding into, and a public Anthropic will intensify it. Interviewers reportedly ask candidates how they would feel if the stock fell to zero, a test of mission over money. Fortune

Big tech's debt insurance is costliest since 2021. The price of insuring against a debt default at Amazon, Alphabet, Microsoft, Meta and Oracle has reached its highest level since 2021, the Wall Street Journal reports, as investors, traders and executives all buy protection around the AI trade. Credit markets are now pricing genuine distress risk into the companies whose compute your product roadmap depends on. Your AI plan carries a counterparty assumption, and it just got more expensive to hedge. The Journal describes the protection buying as tied to trillions riding on the AI trade. Wall Street Journal (via Techmeme)

Broccoli AI — A startup whose agents answer calls and handle job quotes for plumbing, heating and electrical contractors has landed $22M in an early round and $30M in a follow-on, and a Fortune source pegs its worth near $300M. Paying agent customers are no longer only software teams. If agents can run a plumbing shop's front counter, the same economics apply to your service and operations desks — and to whatever your competitors automate first. Fortune opens its story inside a Sacramento plumbing shop. Fortune (via Techmeme)

Claude Ultra — Anthropic renamed its personal tiers — Max at $100 a month, Ultra at $200 — and removed the five-hour usage cap for verified Ultra buyers, though weekly limits remain. The fine print: Ultra's listed usage is now ten times Pro per week, down from the twenty the old name implied, and a June lawsuit claims Anthropic oversold both tiers. If you buy seats, price what the tier actually delivers this month rather than what its marketing name suggests — the headline allowance quietly shrank while the price stayed put. Both plans now bundle API credits: $100 a month on Max, $200 on Ultra. The Next Web

Proton — The Swiss privacy company's CEO cancelled plans to re-interview nearly all staff about their AI use, days after 404 Media reported he was pushing AI through the company despite security warnings from employees — including an engineer whose request for testing time was mocked in front of colleagues. He apologised at an all-hands and on Reddit, but says the push continues. A case study in how an AI mandate turns into a reputational problem: speed pressure aimed at a privacy firm's own security people, with paying customers watching and cancelling. How you roll AI out matters as much as whether you do. Staff say the same measures will now move into performance reviews, and Proton has denied installing software to measure employees' AI-written code. 404 Media

OpenAI — From December, paid ChatGPT plans share a single token-based allowance across Chat and Codex, with Work features folding into the main app; Pro allowances were doubled on October 29. Your seat maths changes in December — get per-plan token figures before renewal. BleepingComputer

White House keeps its AI-testing draft under wraps. The administration won't produce the full text of its draft rules for voluntary AI testing, invoking secrecy and business-confidentiality privileges — so no outsider can check what a 'voluntarily tested' system was actually tested for. Treat the phrase as marketing until proven otherwise. The Information (via Techmeme)

Data centres are now an election issue. Up to 15 of the 36 governorships decided on Tuesday may be won by candidates who want to halt new data-centre construction, an analysis finds, including races in Texas, Ohio, Michigan and New York. Where compute gets built — and whether — is now a political variable in your expansion planning. The Information (via Techmeme)

Cursor extension follow-up — Microsoft confirmed the tampered Claude Code extension never reached its marketplace or npm, and Anthropic's systems weren't breached — official channels were safe; Cursor's marketplace wasn't. Microsoft Security Blog