We tend to cast AI as a race between the United States and China. The deeper contest, write Hoover Research Fellow Dan Wang and Julian Gewirtz, is between the public power of governments and the private power of frontier AI companies. The bigger threat to America's long-term technological edge, they warn, may be Washington's own missteps. Read more in The New York Times: https://lnkd.in/dXWYfydA

AI agents operate at machine speed. When an agent starts making incorrect decisions, taking actions outside its intended scope, or leveraging privileged access improperly, the damage can spread across systems before humans realize what happened. "In many cases, organizations with rogue agents don’t find out from their own detection tools, but from customers, auditors, or broken downstream systems, he says." https://lnkd.in/eXkm47e6

Kimi K3 is out and it should serve as a wake up call to our AI regulators. Anthropic's Fable the current best model in the world, which has been castrated for fear of how the public will use it now has a competitor that's a stones throw away in intelligence, much cheaper, and on the 27th will be completely open for anyone to use however they wish. You can use Kimi K3 to do all the things that Fable was trying to prevent you from doing. It doesn't have no guardrails, but it's not going to stop you from finding attack vectors in your software, or doing chemistry or biology research like Fable does, it's just going to keep you from talking about Tienanmen square. Worse still, I don't really trust US tech giants with my data, but I sure as hell don't trust Chinese tech giants with it, and until the 27th, if you want to use its new capabilities, some of which are state-of-the-art and better than anything you can get anywhere else, you have to send all your data to China. That is the screwed up world we live in now. Somehow the communists are the western middle class's vanguard against our tech overlords and government from conspiring together to keep the most valuable technology for themselves.

Here is my Claude tip of the day: The Inter-Agent Cryptographic Sign-Off: For enterprise workflows, require each agent to append its unique string signature to the payload, ensuring data hasn't bypassed critical validation checkpoints. Let me know if you found this tip helpful, or if you have a better way to accomplish this. #ClaudeTips #GenerativeAI #AICoding #PromptEngineering

Demis Hassabis wants a US-led watchdog to test frontier AI models before release. That is either responsible statecraft or the cleanest cartel optics in modern tech. Possibly both. The notable part is not that someone asked for guardrails again. It is that top lab leaders are now sketching actual machinery for pre-release oversight instead of doing the usual keynote fog about safety. Once the people building the strongest systems start proposing referees, you should assume the stakes moved. Not in theory. In practice. The industry spent years saying innovation moves too fast for oversight. Now even the builders sound like they would prefer an adult in the room.

Tool migration is costly and painful. Mark D. shares how discovering the AI capabilities of Kernaro within Sparx EA allowed his state government agency to cancel a massive migration entirely.

Treaties do not run on vibes. Akash Wasil, Thomas Reed, Jack William Miller, and Peter Barnett map ten ways to catch unauthorized AI development, from satellites and power use to customs records, whistleblowers, inspections, and chip reporting. Every method has seams, so accountability arrives as a stack.

https://lnkd.in/ex4r5Dg8 Good read, with a little addition: America’s overwhelming AI dominance and the resulting barriers to access to frontier AI and its consequences for „non-aligned“ nations.

Attorneys beware, if you use your computer during deposition, fact finding or examination, and you are connected to the Internet, you are sharing every bit of information you post and putting your clients at risk. Having a local Synthony AI Cub, Will change how you do your job forever.

The SOC finally got a fast brain. Agents that triage ten thousand alerts before your coffee cools. And I keep hearing the same relief in people's voices, like speed was the thing that had been failing us. It wasn't. Judgment was. I spent years in the validation world watching tools swear everything was fine, right up until we tested them. A fast system that is confidently wrong is not progress. It is the old problem, now at machine speed. Let the machine think fast. Keep a human thinking slow. And log every call the agent makes so you can actually go back and check it. Autonomy you cannot replay is not autonomy. It is faith with a budget. Would you let an AI close your worst alert, alone?