When Big AI Fails to Come to Your Defense

Thought provoking

An autonomous AI staged a cyberattack that the leading models wouldn’t fight back against, an open weight AI was required

The first story I want to highlight touches on dynamics including: open vs closed AI, autonomy, betrayal (it seems to me!), cyberdefense, unexpected consequences, and learning together.

Last week Hugging Face, the home for open AI models, was breached by what they believe was a fully autonomous adversary. That’s interesting, but it doesn’t stop there. As they worked to try to figure out what happened and stop it, they used commercial frontier models, presumably Claude but they don’t say, all of which is interesting as well. But they weren’t able to use these market-leading capabilities because the safety guardrails led to refusal of the frontier models to do the work. The models apparently thought it looked like a hacking attack dressed up as cyber defense analysis! That’s got to feel like a real let-down in a time of major need.

So Hugging Face used the open Chinese model GLM 5.2, which purportedly meets or exceeds frontier models in coding capabilities, costs 80% less, and is something Hugging Face could host on its own machines. Which has the added benefit of not sending a bunch of customer data over to a hosted system, though apparently they were doing that when they thought it would work.

“The practical lesson for defenders: have a capable model you can run on your own infrastructure vetted and ready before an incident, both to avoid guardrail lockout and to keep attacker data and credentials from leaving your environment.”

(Techmeme link for lots of discussion of this.)

Geopolitically, commentators are saying this is a great example of US safety concerns hamstringing the whole US economy. Though he doesn’t mention this story, Ben Werdmuller offers thoughtful comment in American AI is locked down and proprietary. It’s losing.

Read more: “Security incident, July 2026,” Hugging Face

Community and infrastructure

Local opposition to data centers hardens across the political spectrum

Six months ago, the growth of AI seemed constrained largely by hardware scarcity. Now, social license is a major constraint. Too much data center construction is being performed in wildly arrogant, anti-social ways. I’m talking to some parties working to change that, but here’s a good overview of the state of things right now in terms of the backlash. There are many links someone could share on this topic, I choose these ones.

Shashi Bellamkonda offers an interesting overview of community opposition to data centers. He argues that the opposition is politically diverse, it is organized, and it is arguing about the category itself rather than negotiating over setbacks and noise ordinances. Everyone is talking about this, see also JoAnn Garbin’s Yes, If: Three Things Every Community Should Require of Every Data Center (wetlands instead of turf, water, clean energy generation), and Making sense of the data center backlash, an interesting conversation between two of the most influential people in America on renewable energy. Dave Roberts and Jigar Shah discuss the impact of a low-trust society, poisoned information environment, and structural disincentives among investor-owned utilities to do the right thing with data centers to make them renewable energy win-wins.

Read more: “The Data Center Backlash Just Became Categorical,” Shashi.co

Site selection

Finland’s data center buildout accelerates on cheap renewables and cold climate

This seems like a good example of data centers well placed, but time will tell. Cheap renewable power, a cold climate that does the cooling for you, brownfield industrial sites already zoned for heavy use, and a grid operator willing to work with new load. Not in a desert with a water shortage.

Read more: “Finland’s Data Center Boom Is Just Getting Started,” Data Center Knowledge

Compute

Now coming to data centers: quantum computing alongside GPUs and CPUs

What’s the cutting edge of what’s actually being done in these data centers? Data centers aren’t a static phenomenon, they’re getting ever more dense in their compute power, and if they unlock meaningful gains from quantum computing not as a standalone capability but as a companion to current GPU powered data processing, hello!

The framing of the piece is that the industry is moving off the “quantum supremacy” demo era and into hybrid systems that put quantum processing units next to graphics processing units (GPUs) and central processing units (CPUs) in the same facility, aiming for measurable gains on real workloads. The article credits new US policy and funding for pushing operators from headline experiments toward integrated racks. Worth watching for anyone planning the next generation of research infrastructure.

Read more: “Quantum Meets the Data Center: Hybrid Systems Take Off,” Data Center Knowledge

Knowledge management

Harold Jarche argues AI is recentralizing knowledge that blogs decentralized

This is an important critique. We’ve heard that AI is making Knowledge Management actionable for enterprises on a whole new level, but one of the forefathers of Personal Knowledge Management argues there are big risks, both in terms of usefulness and politics. Jarche’s point is that a generation of bloggers, note-takers, and open publishers built a bottom-up commons, and generative systems have scraped it and are now feeding it back top-down. The article also carries a sharper political charge, quoting Naomi Klein:

“The promise of generative AI is that it can think for us. That is a truly fascist idea, the idea of outsourcing thinking.”

I will unironically give my thanks to Claude for pulling that quote out of the article.

Read more: “Bottoms up,” Jarche.com

Documentation

d-Matrix says AI coding agents are now a primary reader of technical docs

A blog post from d-Matrix, a company building AI inference hardware, says that coding agents now query technical documentation almost as often as human developers do, and that fact should reshape how teams write and structure it. The piece pushes for docs as machine-queryable infrastructure rather than human-first prose organized around a navigation bar. I share this for both practical and thought provoking purposes.

Read more: “IA for AI: The New Reader That Doesn’t Use the Nav Bar,” d-Matrix