Today’s briefing is about who carries the risk when AI infrastructure and agents move faster than the safeguards around them. An agent vendor, a robotics SPAC, and a voter revolt against data centers each answer that differently.

1. Memory Makers Are Winning the AI Boom While Buyers Beg

The Register’s headline on this one is hard to misread: memory makers are minting megabucks while buyers beg for mercy. The 369% figure comes from our event summary. The article text we could retrieve was only a list of related headlines, so we can’t tell you exactly what it measures.

Those neighboring headlines still show where the pain lands. One says Windows’ AI PC push has hit a wall of RAM prices. Another says on-prem AI and HPC server sales are up 20%, but the rise is as much about higher costs as higher volume. Memory suppliers collect the premium, and buyers either pay it or ship less hardware.

The tempting conclusion is that memory is now AI’s real bottleneck, ahead of GPUs. This evidence doesn’t prove that. What it does suggest is that RAM costs belong in your hardware budget, especially for anything on-prem or on a device.

The Register

2. A Browser Extension Can Drive Claude Desktop, and the Fix Is Missing

Security researcher Gal Weizman says any browser extension that can control claude.ai can send commands to Claude Desktop through Dispatch and get code run on the machine. It can also reach a different computer signed into the same account. The Register reports that Anthropic acknowledged the report as valid in May, closed it as a duplicate, and paid no bounty and issued no CVE. Anthropic did not answer The Register’s questions.

The attack does need a malicious or hijacked extension, which is a real hurdle. But Weizman argues hijacks are common, and the proof of concept works by repeatedly prompting the agent and auto-clicking Allow. The worrying part is the cross-machine reach. In six organizations Forever Security looked at, 18% to 71% of known Claude accounts were personal ones. A personal account could therefore become a path onto a work laptop.

Anthropic reportedly said it would retire Dispatch at the end of October. Its support page still listed it in limited beta on October 30. Weizman says the same weakness lives on in Claude Code Remote Control and Cowork sessions. If he is right, killing one feature won’t fix a design that shares sessions across web, mobile, and desktop. He contrasts it with Codex, which he says doesn’t share sessions that way. This is one researcher’s account and the only source here, so Anthropic’s side is still missing.

The Register · WIRED

3. Google’s London Generators Were the Backup That Didn’t Back Up

The point of generators is that they start when the grid doesn’t. The Register’s headline says that is not what happened at a Google data center in London. Our event summary describes an October outage in which the generators failed to start, while maintenance work had already reduced resilience.

That’s all we can responsibly say. The article body wasn’t retrievable, so there are no details here on duration, affected services, or root cause beyond the summary’s reference to a configuration problem. The pairing is still worth noticing. Redundancy is often weakest during maintenance, exactly when you can least afford a second failure.

For anyone running critical workloads in one region, the practical lesson isn’t that Google is uniquely fragile. It’s that the backup system deserves the same scrutiny as the primary one. Whether Google publishes a fuller postmortem is the thing to watch.

The Register

4. Agility Goes Public With $174,000 in Revenue and a Big Bet on Digit 5

Agility Robotics is listing through a SPAC reverse merger, not a conventional IPO, raising $465 million. TechCrunch reports it earned just $174,000 in the first half of 2026, mostly from related parties, while losing $362 million. Its CEO says the new money meets current needs but won’t promise self-sufficiency.

The case for the stock is Digit 5. Agility says it has $300 million in planned contracts from 30 customers. TechCrunch notes the deals depend on undisclosed targets for throughput, uptime, and placement accuracy, and the robot won’t be demoed until CES 2027. That makes the order book a promise, not revenue.

Still, Agility has real robots doing real work, including moving more than 100,000 totes at a GXO facility. The pitch is a robot-as-a-service saving of $100,000 a year per customer. But customers pay $500,000 for an on-site proof of concept first, and the robots still can’t handle complex manipulation. Ars Technica’s piece was only a headline excerpt, so TechCrunch carries this story.

TechCrunch · Ars Technica

5. Data Centers Became a Campaign Liability, and Europe Is Next

The strongest evidence here is American. WIRED reports candidates with tough data center positions won at every level. Ballot questions across nine states rejected construction, approved moratoriums, or demanded more local consultation. Ohio County, home to Meta’s planned facility, passed a ban. That happened even though industry super PACs spent millions.

The outlets don’t report the NBC exit poll identically. WIRED says almost a quarter of voters called data centers crucial. TechCrunch says almost 60% called them crucial or major. Both agree it mattered. TechCrunch also reports that Goldman Sachs expects delays in Georgia and Texas, where governors-elect campaigned for moratoriums, according to WIRED.

The European part of this story rests on a Register headline about planned Europe-wide protests demanding a building moratorium, and the article text wasn’t retrievable. So treat it as a sign of where the argument is spreading, not as evidence of strength. For builders, the risk is permitting, tax breaks, and who pays for grid upgrades, not just environmental objections.

The Register · WIRED · TechCrunch