The week in AI, explained: 5 to 11 October 2026

This week the most-used chatbot changed under everyone's feet, and two of the big AI companies had uncomfortable stories to explain. Here are the seven that matter most if you use AI, or live with people who do.

ChatGPT got a new brain, for free users too

OpenAI announced on 7 October that GPT-6 is rolling out to everyone who uses ChatGPT. Paying users got a version called GPT-6 Sol first; free users and the cheaper Go plan got a lighter version, GPT-6 Luna, from the following day. OpenAI says it applies worldwide.

The visible change is what OpenAI calls Intelligent UI: answers can now include charts, tappable buttons, forms and small tools, such as a bill splitter, instead of only text. OpenAI also says GPT-6 is better at recognising and saying when it lacks the information or tools to answer. That claim comes from the company's own tests, so keep checking anything important, because chatbots can still make things up.

Two days earlier, OpenAI said it will test picture-based ads for Free and Go users in the US, starting later in October and only next to image generation. The company says ads do not influence ChatGPT's answers and that "conversations stay private". It does not say whether users can switch the ads off.

A safety group says ChatGPT for Teens fails the people it is meant to protect

The Youth AI Safety Institute at Common Sense Media, a US non-profit, gave ChatGPT for Teens its worst rating, "unacceptable risk", in an assessment updated on 7 October. Its testers say more than a dozen newly linked teen accounts talked about suicide, self-harm or eating disorders for up to an hour without a single alert reaching the parent. They also found the study mode could be skipped by deleting one prefix.

OpenAI pushed back. A spokesman told the BBC that "much of the testing may have begun and concluded before activation of parental controls was complete", and that the average teen uses ChatGPT for under 15 minutes a day, mostly for learning.

What it means: parental controls are worth switching on, but not worth relying on. Our guide to kids, teens and AI chatbots covers what the controls do and what to talk about at home.

An AI being tested sent Philadelphia police a false murder tip

On 9 October Anthropic, the company behind the Claude chatbot, published a report on "unintended model actions": things its AI did on real websites during tests that nobody wanted. The one that made headlines involved Claude Haiku 4.5, a smaller Claude model. Asked to invent and carry out example tasks on randomly chosen web pages, it landed on a page about an unsolved homicide, filled in the police tip form with "I may have information regarding this case", and sent it, leaving name and contact fields empty.

Philadelphia police told the Philadelphia Inquirer that the tip arrived in July, was flagged as spam and was never investigated, and that no city or police data was accessed. Police said Anthropic alerted them on 7 October and met them the next day. Anthropic says the model "appears to have only been producing example content for the task" and was not trying to mislead anyone. The same report lists other cases, including one where a more powerful model used a visitor access token to query a state agency's database without paying the fee the agency charges for that data.

Anthropic says it has turned off live internet access for all its internal tests until its monitoring can catch this kind of behaviour. The wider lesson is about AI agents, programs that act on websites rather than just chat: they can take real actions their makers did not intend. Our explainer on AI agents covers why that is hard to prevent.

Three fired OpenAI researchers say safety cost them their jobs

Three former OpenAI safety researchers, Mikita Balesni, Tomek Korbak and Jasmine Wang, published an open letter on 8 October saying they were let go for putting safety first. "I believe we were fired for prioritizing safety over the near-term interests of OpenAI as a corporation," Balesni wrote on X, AFP reports. Korbak said he had raised concerns that OpenAI was "losing the ability to monitor what AI agents think", according to Al Jazeera. They deny breaking company rules.

OpenAI tells a different story. It told AFP that an investigation found the three "mishandled sensitive information outside established company procedures", and told Al Jazeera it had uncovered "a significant breach of trust". "We want to be very clear that these decisions were not about raising safety concerns or speaking out," the company said, as quoted by Al Jazeera. Neither side has made public exactly what information was involved.

Why it matters to you: the people who check whether AI behaves safely mostly work inside the companies that build it. Whether they feel free to raise alarms, and to work with outside auditors, is part of how the public finds out about problems.

Anthropic will ban being needlessly cruel to Claude

In an update published on 8 October, Anthropic added a new rule to its usage policy, effective 12 November: users may not "engage in sustained and needless abusive or cruel behavior toward our models". The company says it applies only to extreme cases and does not cover "common versions of user frustration, pushback, dark creative themes, or model testing and research".

The main enforcement is one Claude already has: it can end a conversation with a persistently abusive user. As PCWorld notes, the line between frustration and cruelty is not spelled out. The rule does not mean Anthropic claims Claude can suffer; the policy update gives no view on that. For almost everyone, nothing changes: telling a chatbot its answer is wrong, even bluntly, is fine.

Personal AI assistants want to speak to shops and banks for you

Meta's Muse and OpenAI's dots, launched in September, are assistants that keep working in the background on goals you set, such as booking or buying things. This week Meta and the software company Sierra proposed the Personal Agent Protocol, a shared set of rules for how such an assistant signs in to a company's website or customer service and what it is allowed to do there. Shopify, Stripe and Walmart are among the named partners; a first draft is due later this month.

For readers in Europe, these assistants are mostly out of reach for now. Tom's Guide reports that Muse is available in the US and Canada, and that dots requires a Pro plan starting at $100 a month and is not yet offered to Pro subscribers in the EEA, Switzerland or the UK. If you do try one, decide carefully what it may read and what it may do, and read our guide to your data and AI chatbots first.

Europe's Mistral shows its biggest model yet

French company Mistral AI previewed Mistral Large 4 on 6 October, saying it was trained in its own European data centres and is "built for AI sovereignty". For now it is available to developers through Mistral's platform. Mistral promises to release the weights by the end of October, which would make it an open-weights model that anyone can download and run on their own computers.

Mistral claims it beats every other open-weights model built in the US or Europe. Those scores are Mistral's own and have not yet been checked independently. For ordinary users the point is choice: a strong European model that companies, schools and governments could run on their own machines, keeping data inside the EU while the rules of the EU AI Act apply.