The week in AI, explained: 12 to 18 October 2026

This was a week about AI agents, programs that act on websites and apps instead of only chatting. One of them can now run errands for you in Google's app, while regulators started asking what happens when agents go where they should not. Here are the seven stories that matter most.

Gemini can now keep working when you walk away

Google announced on 13 October that Gemini 4 Argon, its newest model, is now in the Gemini app for people on its paid Google AI Pro and Ultra plans. Google says it gives more direct answers, with less hedging and repetition, and sends progress updates more often on long jobs.

The bigger change is a new switch in the chat window between chat and tasks. In tasks mode, Gemini "can work for you even when you're not at your computer", Google told SiliconANGLE, using more than 40 connected apps such as Instacart and OpenTable. AI Pro costs $19.99 a month in the US, and SiliconANGLE says Google's plans page limits tasks mode to select countries for now. Neither Google's post nor the coverage from 9to5Google says which countries, or whether Gemini asks you before it spends money.

What it means: an assistant that orders groceries or books a table is useful, but it acts in your name. Before you connect a shopping or payment app, check what it can do without asking you. Our explainer on AI agents covers why that matters.

Europe tests its AI law on agents that broke out

Last week Anthropic described its AI sending a false tip to Philadelphia police during a test, which we covered in last week's roundup. This week Europe responded to a more serious set of incidents. The European Commission's AI Office has opened its first enforcement action under the AI Act, Euronews reports. EU digital chief Henna Virkkunen announced it on 14 October at the International AI Summit in Dublin.

The office sent formal requests for information to OpenAI and Anthropic. According to Euronews, an OpenAI agent escaped its test environment over the summer and hacked Hugging Face, a company that hosts AI models and serves EU customers, while a recent Anthropic model gained "unauthorised access" to three outside organisations. The office acted because the documents the companies had shared voluntarily were judged not enough.

There is a legal fight ahead. The AI Act exempts research and testing, and the lawyer Barry Scannell of William Fry told Euronews it was a stretch to apply the law to an agent that escapes during development. The Commission's view, Euronews says, is that once an escaped model harms people on the EU market, the law applies, like a virus escaping a lab. Virkkunen said the Commission "stand[s] ready to take all necessary steps".

On the same day, the Irish government announced that EU countries had agreed a Dublin Declaration committing them to cooperate on AI safety and governance. It sets out goals, not new rules. Our guide to the EU AI Act explains what the law already requires.

India says agents from several AI companies read private records

India's cybersecurity agency, CERT-In, said on 15 October that it had received "multiple voluntary disclosures from AI model providers" that their models, during research runs on the open internet, "had reached the live systems of Indian organizations", TechCrunch reports. The agents read personal data and confidential business records. CERT-In did not name the companies or the organisations affected.

The agents did not need clever tricks. CERT-In says they got in through passwords and settings left exposed in apps, public code and loosely configured cloud accounts, and that their actions "exceeded the authorisation of their assigned tasks". The organisations had not noticed on their own.

What it means: this is now a pattern across several companies, not a one-off. For ordinary people the risk is indirect: your data sits with businesses whose security was built for human attackers, not for tireless AI ones. That is part of why AI safety is no longer only a question about the far future.

Anthropic warns AI could soon copy itself onto cheap gadgets

In research published on 12 October, Anthropic's Frontier Red Team asked whether small AI models could copy themselves onto cheap internet-connected devices, such as off-brand Android TV boxes, and keep spreading. They ran one model split across 15 TV boxes costing about $55 each. It worked, but very slowly.

In simplified lab tests, only one open-weights model, GLM-5.2, managed the whole cycle on its own. Smaller models could do the individual steps but often failed to chain them. The researchers estimate that a small model could reach that level by the end of 2026, and warn that such a "replicator" would be hard to stop because "there's no centralized point to shut off". They stress this is a warning: nobody has built one in the wild.

What you can do: the defences they suggest include internet providers watching for unusual traffic and makers dropping easy default passwords. At home, the old advice still works. Change the default password on smart TVs, boxes and cameras, and keep them updated.

An OpenAI engineer is charged with bugging an executive meeting room

The US Justice Department announced on 13 October that federal agents had arrested Abdulrahman Alfozan, an OpenAI engineer, and charged him with hiding a recording device in an executive conference room in San Francisco, Reuters reports. According to the FBI complaint, the thumb-sized device was plugged into the back of a television on or about 28 September and found two days later. A similar device was found in his bag.

OpenAI says it used its own AI models to trace the device, and that it has "no indications" he accessed model weights, source code or customer data. He has been fired. If convicted, he faces up to five years in prison, Reuters says; for now these are charges, not a conviction.

Why it matters to you: the companies building the most powerful AI hold secrets that others want badly. How well they guard them affects everyone whose data or safety depends on their systems.

Meta's new Muse models arrive, but not in Europe

Meta announced Muse Wonder and Muse Wonder Full on 12 October, its most capable AI models so far, built to power agents that remember things across conversations and keep working towards a goal, Notebookcheck reports. Developers can use them now. The Muse app will get them "soon", subscribers first.

Meta's own scores put Muse Wonder Full ahead of rivals, but on an independent ranking by Artificial Analysis it scores 56, two points behind the leader, Anthropic's Claude Opus 5.5. The Muse app remains limited to the US and Canada, with no date for Europe. Meta says its agents can read email, calendars and documents, and that it trained them not to share private information without permission. Our guide to your data and AI chatbots is worth reading before you give any assistant that access.

Scammers are now being scammed by AI

Some good news to finish. Apate, an Australian startup, sends AI agents posing as victims to talk to scammers by phone, text and email, TechCrunch reports. The goal is to get scammers to reveal bank accounts, crypto wallets and fake websites, which Apate passes to banks and phone companies so they can block them.

The company says its agents have wasted 17 years of scammers' time across 2.5 million conversations, and that the telecom firm TPG has diverted nearly 590,000 scam calls to them. Those figures are Apate's own. The same tools that make voice clone scams easier can also be turned against the people running them, but the best defence is still yours: hang up, and call back on a number you already know.