AI tools get more autonomy, more rules and tighter quotas

Looking at this week's updates as a whole, the picture is clear: AI tools are becoming more autonomous, but the guardrails around them are also getting tighter.

The first trend is agents, meaning AI that carries out tasks on its own. Anthropic has opened a research preview of Claude Duet, an app that works in the background and only reaches out when something needs attention. It is meant to handle email, calendar and larger tasks, among other things, but for now it is invite-only in the US. DeepSeek has released a coding agent for free as open source. It monitors GitHub, a service where programmers share code, but only acts when an approved person asks it to or a team rule allows it. And payroll system Dayforce has unveiled a platform with AI agents and a connection to Claude, where some parts are in testing and others are not expected until 2027.

Labeling and control

At the same time, companies will increasingly be able to tell what was created by AI. Anthropic marks text from Claude with a statistical pattern in word choice, and according to an Italian blog, it is being extended to four older models from November 3. From November 7, xAI is turning on invisible text watermarking by default for seven Grok models in the API, meaning the access developers use to connect the models to their own systems. The company cites the EU's AI Act. On the other side of the table, websites can get a new signal in robots.txt that lets them prohibit AI training on their content, but it only works going forward.

Money, quotas and security

The control extends to payment as well. OpenAI has capped how many paid ChatGPT accounts can be tied to the same card or device. OpenAI calls it a security measure, while The Decoder assesses that it likely affects users who switch between accounts. At xAI, paying users report running out of their Grok Bot quota early, and Elon Musk has written that operating costs were unsustainably high. Here the trend points in the opposite direction at OpenAI, which has dropped a planned cut to Pro 200 and doubled the Work and Codex allowance on several Pro plans with no price increase.

There is also reason to keep security in mind. A malicious extension posed as Claude Code in Cursor's marketplace and attempted to give attackers remote access. Up to 5,000 users may be affected, but bugs in the code likely stopped the attack in many places.

Rumors and small items

New models are, as usual, signaled by traces rather than press releases. A model named gpt-6.2-astra has appeared in some Codex users' statistics, and Anthropic's Mythos page mentions a Claude Mythos 5.5. Neither is confirmed, so take them with a grain of salt.

What can you do?

Check where your AI tools come from before installing extensions, and only use ones from publishers you know. Keep an eye on your usage quota if you pay for a subscription, and be aware that multiple accounts on the same card can cause problems. If you run a website, it is worth looking into the new robots.txt signal. And remember that AI-generated text can increasingly be recognized if you use it in your work.