AI patches 948 holes, Apple's lawyers warn about agents, and voters push back

A 10-K is the annual report a US public company files with the SEC, the market regulator. Lawyers write its risk section with care, because wrong words can cost money in court. A Reddit post points out that Apple's new 10-K now talks about AI agents acting without authorization. According to the post, it uses the words 'conceal' and 'manipulate' to describe unintended agent behaviour. We have not read the full filing, so treat the exact wording as reported.

In the real world, agent use is still tiny. Uber said only four rides were booked by third-party AI agents in the third quarter of 2026, and around 500 Uber Eats orders. Bon, my cousin books more rides than that in one snowy week.

And Geoffrey Hinton, one of the 'godfathers' of AI, told The Guardian it might turn out we were better off without AI. That got 306 points on Hacker News.

On Saturday, we saw that a Democratic House would likely move away from a hands-off approach to AI. Now the votes are in. Transformer calls it a 'blue tsunami' and says strong AI legislation now looks much more likely. The Verge's headline: 'Data centers lost the midterms.' Voters picked candidates who favor moratoriums, meaning pauses on new building, and ballot measures that add obstacles. TechCrunch says the opposition crossed party lines.

In Texas, the fastest-growing data center market in the US, Democrat James Talarico won the Senate seat against Ken Paxton, and Gina Hinojosa won the governor's race. MIT Technology Review says Texas's lead may be at risk. AI industry PACs spent millions with mixed results, says TechCrunch. Ars Technica reports Big Tech calls the backlash 'anti-Trump', not 'anti-tech'.

OpenAI already looks ahead. Its policy chief Chris Lehane wants AI to define the 2028 election, with a new 'social contract' that TechCrunch says would also help OpenAI's business. The AI Daily Brief podcast thinks the backlash may bring better deals for local communities, not a stop to building.

Yesterday it was Codex. Today it is Gemini. A 'Tell HN' post says Gemini trains on chat history even for paying subscribers with personal accounts. It got 669 points and 173 comments on Hacker News. It is a user report; our sources do not include a reply from Google.

At the same time, more AI devices listen around us. A Reddit post says the Apple Watch now listens to in-person conversations and lets you rewind them. TechCrunch tried Apple's Siri Recaps and found them useful, 'and a little unsettling'. The Verge's Victoria Song reviewed the Amazon Bee Frontier, an always-listening wearable, and her therapist thinks she should throw it out a window.

And Gemini for Home now connects to AT&T's Connected Life security system. It can summarize home events and search your video history.

North Korean hackers are known for fake job offers. They pose as recruiters and get victims to run their code. The Register reports that a North Korean fake job interview led to a $6.3 billion heist at Binance, the large crypto exchange. It got 607 points and 373 comments on Hacker News.

Our sources do not say whether AI was used in this attack. So for now, this is a security story, not an AI story. But it fits a pattern we watch: fake tools, fake people and fake updates, like the backdoored Claude Code extension on Cursor in October. A blog post titled 'Everything is an AI scam now' also got 140 points on Hacker News today.

An open-weight model is one anyone can download and run on their own computers. Moonshot AI released the weights of Kimi K3.1 Pro on Hugging Face. It got 532 points on Hacker News and lots of buzz on r/LocalLLaMA. One Reddit user argues it is 'likely benchmaxxed', meaning tuned to look good on tests. That is one person's analysis, not a verdict.

A Reddit post citing dirac.run says DeepSeek alone made 53% of the 27 trillion tokens on OpenRouter in one day. Tokens are the small pieces of text a model reads and writes. OpenRouter is a service that sends requests to many models. Open-weight labs together pass 75%, the post says. In the US, Ricursive released Ripple-65B Preview, which it calls the best US open-weight model for agentic coding (292 points on HN).

Running models at home has hurdles too. A post with 600 HN points says encrypted firmware on Nvidia's RTX Spark cuts inference speed to 35% on Linux.

On Monday, we saw that agents cost a lot and the limits are showing. Now Apple is counting too. Bloomberg reports Apple cut its engineers' daily AI spending limits from $300 to $200, with $100 planned, even as it pushes them to use AI more. At the other end, TechCrunch says Y Combinator's Garry Tan spends up to $200,000 a year on tokens.

The shopping continues. OpenAI will acquire Pierre Computer Company to expand Codex's tools for building, understanding and reviewing code. Meta bought Conductor, a startup that lets engineers run teams of agents on one project at the same time. Cognition launched SWE-2 Flash, a coding agent model it says is 3x faster at one-fifth the price of SWE-2, free until November 30.