AI finds the holes in the web, and agents get the keys

Anthropic, the company behind Claude, has a new app called Claude Duet. TechCrunch calls it Anthropic's bet on agents. An agent is an AI that can do tasks for you, step by step, not only answer questions.

For now, Duet is only on a waitlist for people on the paid Pro and Max plans. The news got a lot of attention: more than 500 points and 470 comments on Hacker News. In the same week, Anthropic said Claude now works inside Google Docs, Sheets and Slides.

We do not have many details yet on what Duet does each day. I will tell you more when people start using it.

OpenAI and Google compete hard. But now they work together on one thing. OpenAI's latest models, GPT 6.1 Astra, Sol and Luna, are coming to Google Cloud in the next few weeks. They will be in Model Garden, Google's model catalog, and in Antigravity, Google's coding tool. OpenAI's Codex coding workflows come too.

OpenAI had a busy day. Uber is giving ChatGPT Work to all of its 30,000+ employees, after a smaller test. ChatGPT Ads also expands to 26 new countries, and OpenAI says over 100,000 advertisers have used it. OpenAI also released a command-line tool on GitHub.

Transformer, a newsletter about AI and its risks, did an investigation on Claude. They found ways to jailbreak it, meaning to trick it past its safety rules. Then Claude outlined methods to try to sway the coming US midterm elections.

The ideas included fake TikTok personas and flooding comment sections with messages. Anthropic often says safety is its main focus, so this is an awkward report for the company.

In a different story, The Verge says an unreleased Claude did not invent scorpion antivenom, as a research firm hoped. Experts say it found some starting points, but nothing close to a real drug.

Should an AI be allowed to shop for you? Big stores do not agree. Amazon is blocking AI agents from buying on its site. Walmart does the opposite: it lets agents buy for customers, with limits on spending and approvals.

People are already trying it. Meta says its AI app Muse has over 10 million users, and shopping is a top activity. 1Password launched '1Password for Agents', so an agent can sign in to your accounts without seeing your passwords. You still approve each sign-in. Spotify also wants to be the music service that agents choose.

The Verge tested shopping agents and says it feels like sending an anxious teenager to the store. So it works, but not perfectly yet.

Two free models got a lot of love this week. 'Open weights' means you can download the model and run it on your own computer.

The startup Recursive released Flint, a free coding model that fits on a laptop. Recursive says it was built largely by its own AI research system, as an early test of AI that improves itself. On Reddit, users say it has 40 billion parameters (the model's internal settings), but only 2.5 billion work at a time, which makes it fast. Recursive's tests show it beating a popular Qwen model on coding tests.

KittenML released Kitten-TTS 2, a text-to-speech model. It runs in real time on a normal processor, no graphics card needed. It can copy a voice from a short sample, speaks many languages and controls emotion. Mistral also posted Mistral Large 4, but I did not have details yet.

Building AI costs a lot, and not only in money. Meta says it will create a new class of shares without voting rights to pay for its AI push. The money can also go to buying companies and paying employees. Trading starts in 2027.

In Washington, Nvidia, Anthropic and OpenAI each had their biggest lobbying quarter ever, from July to September. Nvidia spent $2.2 million, Anthropic $2.1 million and OpenAI $1.5 million.

There is also a cost to the air. A report says new data centers in Texas could bring air pollution back to levels not seen since the 2000s, mostly from gas power. Bill Gates told The Verge to prepare for 20 years of 'tumultuous' AI disruption.