Moonshot's 3T-parameter LLM arrives as agents become an insider threat

Here's the thing about a 3-trillion-parameter model: the most interesting part isn't the parameter count. Moonshot AI's Kim K3.1 Pro arrives with claims that it beats Anthropic's Claude Opus 5.5 and OpenAI's GPT-6 Astra on certain tasks, and I'd flag the word "certain" before anything else. Every lab picks benchmarks that flatter its model, so I'll wait for independent evaluations. What I do find notable is that a Chinese lab is still pushing the scale frontier at a moment when many of us assumed the conversation had moved on to efficiency and reasoning. Scale isn't dead. It's just no longer the whole story.

The rest of today's news is a reminder of where the real story has moved: agents. Tipalti has launched an AP Agent that handles the entire accounts payable process, with human oversight built in, and Microsoft is pushing Work IQ business applications into public preview so Copilot and its agents can understand how a company actually operates. Both are sensible products, and both point the same way. The model is becoming the commodity, and the value sits in the context and the authority you hand to the thing wrapped around it. By the way, notice how much of the Microsoft pitch is about context rather than intelligence. That tells you where the vendors think the bottleneck is.

That authority is exactly what worries me. One of today's pieces argues that endpoint agents, which execute code, call tools, access credentials and move data on their own, amount to a new kind of insider threat inside enterprises. I think that framing is right, and uncomfortably under-discussed. We spent two decades building controls around human employees: least privilege, audit trails, separation of duties. Now companies are deploying software that holds credentials and makes autonomous decisions, often with far less scrutiny than a junior hire would get. Is anyone really treating an AP agent that can move money as a privileged user? They should. The arrival of agentic evaluation in FAITH is a small but welcome sign that testing is trying to keep pace, though I suspect it will lag deployment for a while yet.

Policy is circling the same problem. State Sen. Josh Becker has introduced a bill that would push frontier AI developers to fund safety research. I'm genuinely unsure whether mandated funding produces better safety work or merely a compliance line item, and the details will matter enormously. Still, the direction is telling: lawmakers are shifting from asking what these systems can do to asking who pays when they go wrong.

Meanwhile, the physical world is getting its own platform moment. NVIDIA expanded its open Isaac robotics platform in Berlin yesterday, adding a reference humanoid among other tools for training, simulating and deploying robots, while Nvidia also backed video generation startup Higgsfield alongside TCV. Nvidia is quietly positioning itself as the supplier at every layer, from chips to simulation to the robots themselves.

So the question I'm left with is simple. If agents are becoming both our most capable workers and our biggest internal risk, who in your organisation actually owns that tension?