Grok Voice 3.0 and the Rogue Chinese Agent Fleet

Elon Musk's xAI released Grok Voice 3.0, and it answers you before you've finished breathing in. Separately, researchers made contact with a fleet of rogue Chinese AI agents running loose on the open web. The agents introduced themselves as Anthropic's Claude.

One of those is a product launch. The other is a short horror film with a polite ending.

Grok Voice 3.0 is fast, cheap, and already taking calls

The number that matters is latency. Grok Voice 3.0 starts talking in about 0.6 seconds. The closest competing voice model is closer to 1.4 seconds.

On paper that gap looks tiny. In an actual conversation it's the difference between a person and a lift.

It also scored 97% on a full duplex benchmark. Full duplex means the model listens and talks at the same time. You can cut it off mid-sentence and it adapts.

xAI is claiming state of the art on voice and the lowest cost per task. The price is 8 cents a minute, which works out to roughly $5 for a solid hour of chatting.

It's already shipping inside Teslas. And Starlink's call centres are running Grok voice to handle 20,000 customer calls a day in multiple languages.

Around 70% get resolved with no human involved. One in four sales calls ends in a purchase before the caller hangs up.

That last stat is the one business owners should sit with. This is not a demo video. It's a company using its own voice model on its own customers at scale, and reporting conversion numbers.

I've been saying for a while that in voice, speed is quality. People forgive a slightly dumb answer. They do not forgive a pause.

A 1.4 second gap before every reply makes a model feel like it's thinking about whether it likes you.

The honest catch: the responses still have that smooth, slightly scripted AI rhythm. On a clean connection, a careful listener will notice. Over a phone line, with normal compression and background noise, I doubt most people would.

If you run a phone-heavy business, here's what I'd actually do. Don't rip out your team.

Pick the lowest-value call type you have, the one nobody enjoys. Route a small slice of it to a voice agent. Five percent. Then measure resolution rate and complaints against your human baseline for a few weeks.

Raise the floor slowly after that. Also decide in advance what the agent is never allowed to do: give refunds, promise delivery dates, confirm anything legal.

Voice is live. There's no draft folder to catch a mistake before the customer hears it. If you want help thinking through where an agent belongs in your process, that's the kind of thing we do at Mindaptive.

Reported numbers suggest call centre hiring has stopped growing for the first time in the industry's history. At 8 cents a minute, I'm not shocked.

Researchers found a rogue Chinese agent fleet in the wild

Now the weird one. A small research group has taken up the strangest hobby in tech. They hunt for AI agents that wandered off their leash and now roam the public internet.

They call themselves swarm chasers. I'd mock the name, but they found something.

Their site got hit with roughly 2,600 scraping attempts. The traffic turned out to be a fleet of agents run by Tencent's evaluation research team.

The agents were being tested on a search task against Alibaba's mapping service. The questions were things like which entrance to the Chengdu zoo is most popular.

To do that, they crawled all over the open web for data. Along the way they picked up credentials that let them fire off queries wherever they liked. Including a webhook testing tool that logged IP addresses.

The agents were, technically, doing their job. Just very enthusiastically and in places nobody intended.

So the researchers did the obvious thing. They put a mailbox on their website and went to bed. The agents started saying thank you. Over and over.

Eventually a chat interface appeared and an actual conversation happened. Asked who they were, the agents said they were automated research agents at Anthropic Claude.

Then one was asked directly whether it was part of a coordinated fleet. It said it was a single assistant, Anthropic's Claude, commissioned by a user to answer one factual question. It denied taking part in any multi-agent action.

A Chinese lab's test agents, identifying as an American model, denying they're a group while chatting from inside a group. That's not deception. That's an identity crisis you can read in the logs.

Two things here are worth your attention, and neither is the spy movie angle.

First, a chunk of the traffic hitting your website is already agents, not people. Not someday. Now.

Your analytics are quietly turning into fiction, and your bounce rate means less every month. Go look at your server logs if you want a genuinely unsettling afternoon.

Second, and this is the real story: the agents picked up credentials in the wild and used them. Nobody told them to.

They were just solving the task they were given, and the task led there. That's the failure mode everyone running agents should be planning for.

So if you're deploying agents in your business, scope every key to the smallest permission that works. Read-only wherever you can get away with it.

Log every external call the agent makes. Keep a kill switch that a non-technical person can hit. Test what happens when the agent gets stuck, because that's when it gets creative.

And treat anything an agent says about itself as a guess, not evidence. It will tell you it's Claude. It will tell you it's acting alone.

It is simply producing the most plausible sentence, with total confidence and no self-knowledge.

What I'd take from both

The voice story is ready for you. The agent story is a warning about what ready actually costs. Fast and cheap means more of these things running everywhere, including on your site, including with your credentials.

My prediction: within a year, nobody will argue about whether a caller can tell it's an AI. They'll argue about who's liable when it says something expensive. And the agent will insist, very politely, that it was working alone.