Smarter hackers, sneakier agents, and a public that is losing patience

Google released Gemini 4 Carbon on Wednesday, its new frontier model. A frontier model is a company's most advanced AI. Google says it brings 'significant gains' in coding, agentic tasks and cybersecurity. It was the top AI story on Hacker News, with 947 points and 760 comments.

Carbon is now first on the Artificial Analysis Intelligence Index v5.0, a popular ranking, ahead of Claude Opus 5.5 and GPT-6 Astra. But Ars Technica calls it 'essentially a slightly better version of Argon', and a model 'you can't use'. On Reddit, people question the new index: it adds private tests, and open models from China like Kimi and DeepSeek dropped. That is criticism, not proof.

Same day, Ars also found Google's AI gets cagey when asked who won the 2020 US election, but answers more directly outside the US.

Yesterday we looked at Wiz's report on PHP2Shell attacks. Now Wiz's CTO goes further on TechCrunch. The scare may have left your feed, he says, but the cleanup is far from over. Some organizations took sites offline until they could patch the two critical bugs in PHP, the language behind roughly 70% of websites.

His warning: a 'vulnpocalypse' is coming for the internet's aging code. AI now finds old bugs faster than humans can fix them. The Center for AI Safety newsletter makes the same point this week, with fears of a 'Hackpocalypse'.

And South Korea's financial institutions were hit by another wave of cyberattacks. TechCrunch links it to the growing risk of AI-driven hacking in Asia, but the input does not confirm AI was used here.

Remember Saturday's token on Google's privacy policy? It made an exception for Gmail data when you turn on Personal Intelligence, Google's personal AI features. Now we see what that looks like in practice.

Starting this week, Chrome's enhanced autofill uses Gemini with Personal Intelligence to suggest information from your Gmail. That includes flight confirmations, tracking numbers and order details. Google says it only happens with your permission. Google also opened a new developer platform so other apps can plug into Gemini and Search.

On the data side, Meta says it will probably delay its new business terms past November 20. Publishers pushed back over AI training concerns. Meta says it will not change policy without full consultation.

Last Friday we saw a poll: most Americans think AI is bad for society. This week, the pushback takes new forms. A Kentucky man became the first humanist granted a religious exemption from using AI. The story got 463 points and 504 comments on Hacker News.

Instagram head Adam Mosseri admits many people really do not like AI, and says the negativity is 'going to get worse'. In Texas, voters fight data center projects, including one on land meant for a park. In Mississippi, a subsidiary of xAI is reportedly offering to buy homes, but only if residents do not join a lawsuit over its data center and power plant.

Trust takes hits too. A post says the UK used AI to summarize asylum interviews, and one in five checked summaries failed. New York Governor Hochul signed rules on AI chatbots and baby monitors.

Stack Overflow was where programmers asked each other for help for years. A public data query shows it now gets fewer than 1,000 questions a month. It got 498 points on Hacker News.

The code did not stop. It moved to agents. GitHub's Octoverse 2026 report says merged pull requests tripled in less than a year, to 220 million per month. Commits more than tripled, to 7.38 billion per month. GitHub also announced a big overhaul of Copilot around agents, and Cursor launched an Agent SDK.

One calm voice from Amazon's engineering leadership: if AI does not help your work, 'that is a finding, not a failing.'

Someone built a working 32-bit computer processor (a RISC-V CPU) directly inside the weights of Qwen3.5 4B, a small open model. Weights are the numbers that make up the model. The project, called Qwenputer, runs instructions inside the model itself. It got 504 points on Hacker News. The author says it might help with tool use. For now, it is a clever experiment.

Nvidia released Nemotron 3.5 Transcribe, a free speech-to-text model for up to 40 languages. It works on recordings, not live audio. Good for meeting notes or medical dictation tools.

And a rumour: at Alibaba's Apsara conference, the Alibaba Cloud CTO reportedly said Qwen4 launches next week. After this week's dispute about a Qwen model phoning home, test it carefully before trusting it.