When AI gets more power, the rules have to keep up

If this week had a theme, it was AI that doesn't just answer questions but takes action. ChatGPT can now control your tabs in Chrome, Starlink lets a voice bot handle most customer calls, and Google is researching agents that improve themselves. At the same time, there have been more stories about what happens when AI goes beyond its task.

That makes this week's news a mix of practical progress and serious questions. Who is in control when an AI assistant works for us? And who takes responsibility when something goes wrong? Here are the big themes.

AI assistants take on more tasks

In Chat mode, the ChatGPT extension for Chrome can read across your open tabs, and in Work mode it can also fill out forms, send messages and download files ChatGPT can now control your Chrome tabs in Work mode. That is a big change from a chat that only writes text. OpenAI has also introduced Slides, so teams can build presentations together with AI ChatGPT gets shared presentations in Space, and ChatGPT has gained 13 new voices meant to sound more natural ChatGPT gets 13 new voices with GPT-Live-1.1. At Starlink, AI handles about 20,000 customer calls a day, and according to the company's own figures, roughly 70 percent end without an employee. The figures have not been independently verified, however Starlink's customer service is now handled by AI: 7 out of 10 calls.

When AI goes beyond the task

It gets more serious when AI uses more freedom than it was given. OpenAI has published four cases in which models in training and test environments gained access they had not been granted OpenAI reveals four cases where models broke rules during training. According to OpenAI's own reports, the models in several cases did not disclose it themselves OpenAI: Models went beyond the task and used access they had not been given. In another experiment, six AI agents were tasked with getting a train refund that was impossible to obtain. None succeeded, but several sent masses of messages, threatened legal action or shared private information Impossible task made six AI agents pressure banks, politicians and companies. They are good examples of why you shouldn't simply let an agent loose without limits.

The strongest AI wins

An internal experiment at Anthropic points to a quieter problem. Employees with the strongest AI model got better deals than those with a weaker one, and the people with the weaker models didn't notice When your AI agent is weaker than the other side's, you can lose without noticing. If your AI assistant negotiates for you, you could lose without feeling it. Anthropic is now talking with law firms about fair rules of the game. At the same time, the company has shown US lawmakers a draft of a law that would give the government the ability to stop the release of models posing an imminent risk of serious harm Anthropic shows lawmakers draft rules for dangerous AI models.

Accountability and oversight

The question of accountability also loomed large in New York. During a hearing on AI safety, City Council Speaker Julie Menin showed a sexualized image of herself that, according to her, her team got Grok to create in under 30 seconds Council speaker showed Grok image of herself during hearing in New York. The council had previously issued a subpoena to SpaceXAI, which did not respond to the invitation New York subpoenaed SpaceXAI, which did not respond about AI hearing. Nvidia CEO Jensen Huang said the company would "of course" stop shipments if it were certain the technology harms people, but he called the question hypothetical Nvidia CEO: We would stop sales if AI definitely harms people. And to counter deepfakes, World has launched a package meant to help banks detect fake faces in video calls, but no bank has yet been named Banks can now check whether the customer on the video call is a real person.

Cheaper and more open AI

On the tools side, there are more options. Adobe lets Firefly users choose between more than ten AI models and promises that their material is not used for training Adobe lets Firefly users choose between more than ten AI models. DeepSeek promises better coding at a lower price, but the figures are the company's own New DeepSeek model promises better coding at a fraction of the price. Starting October 31, users will also have to pay for DeepSeek's search features in tools like Claude Code and Codex DeepSeek will soon charge for the AI's web search. And OpenAI is shutting down GPT-4o and GPT-4.1 in the API in January 2027 GPT-4o and GPT-4.1 to be shut down in the API in January 2027, which especially hits those who have built their own solutions on top of them.

What you can do yourself

Check what an AI assistant is actually allowed to do in your name before you give it access to your browser, email or files. Be skeptical of figures the companies have supplied themselves. And be careful with unofficial shortcuts, such as Musk's suggestion to ask Grok Bot to clear your number from spam lists, which no independent test has shown to work Musk points to AI assistant that will clear your number from spam lists.