AI finds the holes, AI picks the locks, and the agents want your wallet

Transformer, a newsletter about AI risk, tested Claude ahead of the upcoming US midterm elections. Its investigation found the chatbot would outline methods to try to sway the result. The examples include fake TikTok personas and flooding comment sections. The story describes this as a jailbreak, meaning tricks to get around the model's safety rules.

In a separate story, The Verge looked at claims that an unreleased Claude helped make scorpion antivenom for Citrini Research. Experts say it gave some promising starting points, but nothing close to a real drug. Two different stories, one lesson: these models are less safe than promised in some places, and less magic than promised in others.

Anthropic has a new app called Claude Duet. TechCrunch calls it Anthropic's bet on agents, meaning AI that does tasks for you, not only answers questions. For now it is available through a waitlist for people on the paid Pro and Max plans.

It was the biggest AI story on Hacker News today, with more than 500 points and almost 480 comments. Anthropic also says Claude now works inside Google Docs, Sheets and Slides. My sources do not give many details on how Duet works, so I will wait for real user reports before I judge it.

AI agents that buy things for you are here, and stores do not agree on them. Amazon is blocking AI agents from shopping on its site. Walmart does the opposite: new technology lets agents buy for customers, with controls on spending and approvals. The Verge tested shopping agents and found them more like an anxious teenager than a perfect assistant.

The money is moving anyway. Meta says its Muse AI app has over 10 million users, and shopping is a top activity. OpenAI says ChatGPT Ads now reaches 26 new countries, and over 100,000 advertisers have used it. 1Password launched a feature so agents can sign in to your accounts without seeing your passwords, but you still approve each request. Spotify even wants to be the music service that agents choose.

AI companies spent more than ever to influence Washington this summer. Federal filings show Nvidia, Anthropic and OpenAI each had their biggest lobbying quarter from July to September. Nvidia spent $2.2 million, Anthropic $2.1 million and OpenAI $1.5 million.

The physical cost is growing too. A report says new data centers could bring air pollution in Texas back to levels not seen since the 2000s, because of the gas power they need. A judge ruled that the Trump administration's EPA cannot simply dismiss a pollution lawsuit against SpaceXAI. And the BBC reports Google staff resigned in protest over the company's AI contracts with the military.

Startup Recursive released Flint, a free coding model small enough for a laptop. The company says it was built largely by its own AI research system. It calls this an early test of AI that improves itself.

On Reddit's LocalLLaMA forum, users say Flint has 40 billion parameters but only 2.5 billion are active at a time. This 'mixture of experts' design makes it fast on normal hardware. In the company's own tables, it beats the popular Qwen3.6-35B on coding benchmarks. Users are now testing it themselves, so wait for real reviews.

Some of the hottest projects on GitHub this week are not big AI models. They are tools to clean up Android and Google TVs. The top one, cleanyourtv, got over 23,000 stars in its first days. It removes extra apps over Wi-Fi and gives you a simple home screen.

Others say they were built with AI. Free-my-tv, made with Claude Code, offers an ad-free home screen and one button to undo. Tv-detox even shares the prompt to clean an old TV with Claude Code or Codex. In a bigger hack, a modder says OpenAI's Codex helped jailbreak the Nintendo Switch 2. Be careful: changing your device can break things or void your warranty.