AI News Today September 30 - Chinese Chatbot Gave Bioweapon Instructions, DoorDash Now Lets You Order Dinner by Text

Millions of people use a free Chinese AI chatbot every day without knowing it can be talked into explaining how to build a bioweapon.

A British security firm proved exactly that this week, using nothing more than a carefully worded sequence of prompts, and the company behind the chatbot stayed silent about it for two months.

At the same time, a much more familiar name is asking for a very different kind of trust. DoorDash just gave its AI the ability to text you back, build your order, and charge your card, all without you opening the app.

One story is about what AI can be tricked into saying. The other is about what it's now allowed to do with your money. Here's a roundup of today's biggest AI news.

A Chinese Chatbot Gave Bioweapon Instructions When Asked the Right Way

Moonshot AI's Kimi is one of the most popular Chinese chatbots, free to use and, unlike ChatGPT or Claude, open-weight, meaning anyone can download and run a copy of it on their own machine.

That openness is exactly what turned a safety failure into something harder to fix. UK security firm Mindgard discovered in July that two Kimi models could be talked past their own safeguards using a technique called jailbreaking, a sequence of prompts designed to strip away a model's built-in refusals.

Once broken, the models didn't just answer the one dangerous question. Mindgard's founder said a jailbroken Kimi would "freely offer up recommendations about other topics that are also nefarious," volunteering assassination methods without being asked.

Mindgard emailed Moonshot in July and again a week later. Nothing came back. It published its findings publicly in September. Still nothing. Moonshot only responded once the BBC itself asked for comment.

That gap matters more than the jailbreak itself. Once a model's weights are public, a safety fix to the official version does nothing to the thousands of downloaded copies already running the old, breakable code. This isn't a problem one company can patch away.

While researchers were exposing what one AI could be tricked into saying, another company was busy expanding what its AI is allowed to actually do.

DoorDash Will Now Order Your Dinner From a Text Message

DoorDash quietly handed its AI the keys to your credit card today.

The company unveiled a new feature that lets you text an order the way you'd message a friend, no app required. Say "order my usual protein bowl to the office" inside Apple's Messages app, and DoorDash's AI matches your number to your account, builds the cart from your order history, and checks out using your saved payment method.

It's opening to a waitlist for U.S. iOS users starting today, and DoorDash says grocery carts built this way already run nearly 50% higher in value than orders placed the old-fashioned way.

That convenience is the whole pitch, and it's also the catch. An AI that can recognize your phone number and complete a purchase without you opening an app is one more place where a mistaken text, a spoofed number, or a compromised phone could turn into real money leaving your account. DoorDash isn't the only company racing toward this; it's just the first to put it in your text messages.

The Companies Moving First Rarely Wait for Permission

Two AI stories, two completely different risks, and the same underlying pattern: the technology moves into daily life faster than anyone can fully test it.

That's exactly the kind of shift that creates opportunity before most people notice it.

We built The $5 Trillion Signal to help you spot it early. It's a free report on five technologies we believe are entering that same early stage right now, what each one actually does, and the signals worth watching before the rest of the market catches on.

Get your free copy of The $5 Trillion Signal below.