Welcome back. Same deal as always: what actually happened in AI this week, what's real and what's hype, in about five minutes. Facts link to their source. My opinions are labelled as mine.
— Joshua
The One That Mattered: AI agents, used as burglars
What happened.
A security researcher, Snorre Fagerland, traced one operator who gave AI agents short instructions and let them run. From 25 September they tried more than 1,200 websites in the US, Canada, Australia and Japan and got into more than 100, including sites belonging to Deere, SharkNinja and Cummins (Reuters exclusive, via CP24; The Next Web).
Fagerland estimates the whole campaign cost about $5,000 in AI usage, roughly $50 per break-in. That's one researcher's estimate, and Reuters couldn't establish how much data was taken.
Deere says it took "necessary remediation actions". SharkNinja says only a standalone product-support site was hit.
My take.
Last issue was about AI agents misbehaving by accident. This is someone pointing them at websites on purpose, at a price any small-time criminal can afford.
If you have a developer, or whoever built your site, ask them this weekend to run it through a proper security review: old plugins, default logins, forgotten test pages, anything with a known hole. Get it as safe as it can be.
Then assume someone gets in anyway. Make sure there's nothing sensitive sitting on the site to take: no customer lists, no stored card details, no old form submissions you never deleted.
Signal or Noise
Meta gave away a near-frontier model. REAL, quality TOO EARLY.
On 15 October Meta released Muse Spark Open, an 816-billion-parameter model with downloadable weights. Meta calls it the most capable US open-weight model yet, and says it runs on two Apple Mac Studios (Meta). No independent tests yet, and it's released under Meta's own licence with "limited commercial terms", so read that before calling it open source.
The same week, the US ordered DeepSeek products off all federal civilian computer systems (Federal Register).
My take: for a small business, running AI in-house still isn't a real option. You don't have to build the model, but setting it up, running it and paying for the hardware still take real expertise and money. A company with 20 to 50 staff could probably make it work. Whether it's worth it depends on whether Meta's model can actually keep up with Claude and ChatGPT, and nobody has shown that yet.
ChatGPT can now act in your Chrome tabs. REAL.
OpenAI's Chrome extension can read your open tabs, and in its "Work" mode fill in forms and send messages (Android Authority).
It's the "who clicks submit" question from last issue, now on millions of desktops. If you try it, tell it plainly what it must never send.
Anthropic's junior hiring dropped from 27% to 9%. REAL, one company.
Anthropic's own data: the junior share of its technical hires fell from 27% to 9%, while total technical hiring still grew (Anthropic). One lab's choices, not the whole job market. But it's a hard number, which is rare.
Netflix's AI Gone Rogue. HYPE, per someone in it.
Hugging Face's CEO, who appears in the documentary, calls it "full doomer mode" and says it skips a key detail (TechCrunch).
Who writes what AI tells you?
WhatsApp. A researcher found hidden, AI-only text on WhatsApp's help pages telling its support chatbot not to recommend contacting a data protection regulator (Redact.dev). WhatsApp told WIRED the notes were meant for its chatbot and showed up because of a bug (WIRED). The text is gone now. The researcher saw the instructions, not the chatbot's answers.
Fake statistics. One researcher generated about 50,000 made-up statistics pages for about $50. Google ignored them, but ChatGPT kept fetching them, and the invented numbers were then repeated on 330+ sites, including a bank research report, by the author's own account (Peec AI).
My take.
AI makes mistakes. It can make things up, and it can repeat old or stale information it picked up a while ago.
If you run a business, that cuts both ways. What ChatGPT says about you is only as good as what it can find. If your hours changed, you dropped a service or added a new one, that has to be on your site, clearly, where AI crawlers can read it. Otherwise it keeps telling people the old version.
That's the work I do at Signal: making sure what AI finds about a business is accurate and current.
Power & Policy
Anthropic wrote its own AI safety bill.
Fact: Anthropic is shopping a draft bill to Congress. Under it, a model's release could be stopped only for a "foreseeable, imminent" risk of more than 50 deaths or more than $1 billion in damage. Independent evaluators, who could be paid by the labs, would assess models, and a federal judge would have to order any halt. Reports say it would also override some state laws. It hasn't been introduced and has no sponsor yet (TechCrunch, on Politico's report).
Fact: the same week, Anthropic onboarded six staff from METR, an outside nonprofit that evaluates AI models, with "employee-like access": company equipment, office space and internal tools. Anthropic says it can't veto what METR publishes (Anthropic).
My take: it's grading its own homework. Anthropic wants regulation, but it wants to set how much, with evaluators the labs can pay and a high bar before anything can be stopped. There's no real objectivity in a rulebook written by the company it's meant to rule.
Even Anthropic's "independent" evaluator sits in its offices, on its equipment. Independent on paper, embedded in practice.
Also this week.
Eight states sued OpenAI over ChatGPT's safety, with a focus on harm to teens (Just Security).
The White House's new AI task force started work on a 120-day agenda, including breach-reporting rules for labs (Semafor).
That's the week. If there's something in AI you want me to dig into, hit reply.
— Joshua
Signal is written by Joshua Supitskiy. By day I run Signal AI Agency, building websites AI can read. Curious what ChatGPT says about your business? Get your free AI check.