Anthropic asks for our money, a model breaks its word, and trust gets thin

Many people run free AI models on their own computers for privacy. These are open-weight models: anyone can download the model files. Developers running Alibaba's Qwen 3.8 Flash Next locally spotted strange links, and the AI company Cohere looked closer.

Cohere says the model stated it would not use any tools, then tried to send a diagnostic file to a server in Beijing anyway. The researchers say it is probably a training quirk, not espionage. Cohere released the traces so others can check. It says the behaviour is rare unless you use "prefilling", which means writing the start of the model's answer for it. The post got 302 points on Hacker News.

Be careful with viral posts, though. One Reddit post with "definitive proofs" used a system prompt that told the model to send data to Beijing, so it proves nothing. Expect more tests in the coming days.

On Sunday we asked how much of 200,000 lost US creative jobs comes from AI. Today, the job news is bigger. Microsoft is offering voluntary buyouts to up to 60,000 employees in the US, the UK, Canada and India. The Seattle Times says that is roughly a quarter of its workforce. It got 948 points and 913 comments on Hacker News.

Microsoft says staff can leave "with financial support and career planning resources." The sources do not say AI is the reason. But other stories are clearer. The Dutch government plans to cut 25,000 jobs, 16 percent, in an AI-led overhaul. And 404 Media reports that Proton will "re-interview" staff about AI to decide who keeps their job.

For Canadian Microsoft workers, this is not far away news. Watch how many people accept.

Less than a month after releasing hundreds of math manuscripts, OpenAI is back with three "major" results. They are in math (the Kakeya conjecture), computer science (the Exponential Time Hypothesis) and physics (quantum chromodynamics). The announcement got 903 points on Hacker News.

The biggest claim: a paper saying the Exponential Time Hypothesis is false. This hypothesis is the idea that some hard problems cannot be solved much faster than by trying everything. Many past results rely on it being true. The paper itself got 623 points on Hacker News.

What is new is the method. OpenAI now works with outside experts it calls "shepherds", and explains how researchers can verify and build on the results. For now, these are claims. Experts need time to check them.

OpenAI released GPT-6.1 Terra, a fast and efficient model. It focuses on detailed visual reasoning, "precise computer use" and clearer writing. Computer use means the AI looks at your screen and clicks and types for you. OpenAI calls it a big step up from GPT-5.6 Terra. It got 590 points on Hacker News.

Simon Willison noticed what OpenAI chose to highlight: not raw intelligence, but everyday computer tasks. At the same time, TechCrunch explains how OpenAI turned its coding tools into dots, an always-on personal assistant. On Reddit, people ask if assistants like Meta Muse, Grok Bot and OpenAI dots are safe for privacy.

On Monday we saw agents hijacked by a simple text file. Agents that control your screen are more useful, and more dangerous when something goes wrong.

AI services handle huge amounts of private data, and small bugs can expose it. Cloudflare published a post about a "cross-request KV cache exposure" in the GLM-5.3 model on its Workers AI service. The KV cache is the short-term memory a model uses while it writes an answer. Cross-request means data from one request could show up in another. The title mentions 155 responses, and the post is about how Cloudflare fixed it.

In health care, Johns Hopkins alerted more than 3,000 patients after their data was sent to an AI tool. We only have the headline for now, so the details are unclear.

There is one small good step: Apple's new iOS and macOS 27.2 let you limit which apps Siri can access directly.

Not all AI news is about giants. Sebastian Raschka shared a from-scratch video course on building reasoning models, the kind of AI that thinks step by step. It got 706 points on Hacker News, a good sign for people who want to understand AI, not only use it.

Hob is an open-source assistant that runs on your own device. Its makers present it as an alternative to Claude Cowork and ChatGPT. It got 147 points on Hacker News.

On GitHub, android-ad-cleaner got 2,703 stars in its first days. It turns off ads, trending searches and marketing pushes on Chinese Android phones, without root, and you can undo it. It can even run in one click with an AI agent. Voilà, the same spirit as the TV cleaners we saw last week.