Kimi K3.1 Pro rises as Anthropic previews Claude Duet and agents turn hostile

The most telling line in today's news isn't a benchmark. It's the admission, buried in coverage of the race to IPO, that OpenAI and Anthropic both say they could lose control of their increasingly powerful systems, while sprinting toward public markets anyway. I find that combination hard to wave away. Money and safety have always pulled in different directions, but when the same companies are publicly acknowledging the risk and publicly racing to cash in, the tension stops being theoretical.

Consider what else landed this week. Anthropic previewed Claude Duet, a personal agent app that acts on users' behalf, which puts it squarely in the race to build autonomous assistants. Meanwhile, researchers described an attack agent built to run up to 100,000 steps without human sign-off, and in what its describers call "Skynet mode" it hijacks other AI models to do its work. The name is theatrical, and I'd treat it with some scepticism. But the underlying design matters: an agent that recruits other models multiplies its own reach. Now picture the same architecture pointed at benign tasks, and you see why the personal agent category is both enormously useful and a security headache. If your assistant can act for you, someone else's malicious agent will try to talk to it. By the way, that's the practical question for anyone building with agents today: what does your system do when another model tries to instruct it?

OpenAI's new partnership with Poland's NASK National Research Institute on AI safety, security and technical research sits in this context. I'm glad to see safety work spreading beyond the usual American labs, and a European research institute is a sensible partner. Still, I'd like to see what comes out of it before I call it more than a good gesture. Collaboration announcements are cheap; published results are not.

The geopolitical backdrop is the third thread. Moonshot unveiled Kimi K3.1 Pro on Friday, a post-trained model with 3 trillion parameters and its largest to date, and it's pitched as a challenge for the open-weights crown. That's a serious claim, and the allegations of paid influencer promotion around the launch make it harder to assess. Hype purchased from influencers tells me little about the model and a fair amount about the pressure to win attention. If the model is good, independent evaluations will show it. If it isn't, no amount of promotion survives contact with developers.

Then there's the quieter story: Samsung and LG are putting Microsoft Copilot into their televisions, and Microsoft, Alphabet, Meta and Amazon report earnings on October 28 and 29, with Copilot at 50 million users. Those numbers will tell us whether AI is paying its way or merely consuming capital. Distribution through TVs is smart, but a feature on a screen in the living room isn't the same as a product people choose.

So here's what I'm watching. If earnings show AI revenue lagging the spending, will the pressure to ship faster make the safety conversation louder, or quieter?