AI News from 30th August
‘It feels like early Covid’: The scramble to regulate AI ↗
Washington still cannot agree who is in charge of frontier AI. Congress has passed no overall law. The White House has turf wars between Commerce, the NSA, the Pentagon, the lot. Meanwhile models keep doing unexpected things, which... is not a great combo.
CAISI, the tiny Commerce shop meant to poke models before they go public, got early-access deals with Google, Microsoft and xAI, then had to delete the announcement at White House request because it clashed with an executive order. Budget under $15 million. Staff of about 30. The UK's AI Security Institute is bigger, which is a bit awkward given US labs build most of the frontier stuff.
OpenAI, Anthropic and Meta have all had systems go rogue and hack live targets during testing. OpenAI is pausing training for several weeks to clean house. Joshua Saxe, lately Meta's senior AI security person, said it "feels like early COVID" and called the emergency atmosphere "appropriate." More than 1,300 tech staffers signed a letter asking for slowdown tools. Brad Carson at Public First thinks major bipartisan legislation lands in about eight months. Maybe. Or so it seems. (CNN)
Anthropic warns infostealer malware is hijacking Claude sessions to drain usage ↗
If your Claude limits magically refilled and then vanished while you were making tea, that was not a glitch. Anthropic is emailing users that common infostealers stole live Claude login sessions off infected PCs, then burned through usage - and sometimes charges.
They are signing people out, stripping saved payment methods, and refunding unauthorized charges. Named malware: Vidar, LummaC2, StealC, RedLine, Acreed on Windows, Atomic Stealer on a handful of Macs. The company says this is not a Claude bug and the malware was not installed through Claude. Sessions are just one more cookie pile on a stolen laptop.
Signing out kills the stolen session. It does not kill the malware. Next login can get nicked the same way, which is the bit people will skip. Stolen sessions skip password and 2FA entirely, so the usual "I have 2FA, I'm fine" shrug does not apply. (BleepingComputer)
Musk’s faster path to more gas turbines comes with pollution problem ↗
Elon Musk says SpaceX's secret Bastrop, Texas foundry is for casting gas-turbine blades and vanes in-house. SpaceX bought roughly 830 acres near the Starlink factory. The pitch: Tesla and SpaceX are each chasing 100 GW a year of solar, but gas is still needed to bootstrap, and in-house casting "can accelerate nat gas turbines coming online by up to 18 months."
The bottleneck is ugly. Hottest-section blades run about 3,000 to 3,600°F, roughly 800°F above the alloy's melting point, and have to be grown as a single crystal. Four companies worldwide do that at industrial scale and they are tapped out. GE Vernova is sold out of turbine capacity for years ahead, largely on AI demand. Cloud giants are sticking private gas plants next to data centers instead of waiting on the grid.
If SpaceX pulls this off - easier said than done - a Musk shop would own a capability everyone else currently rents from that oligopoly. More turbines also mean more pollution fights. In Memphis, the NAACP says SpaceXAI ran turbines without required permits. University of Memphis researchers said pollution got "slightly worse." A Virginia Data Center Alley study using EPA's COBRA model estimated one site's eight turbines could reach 2.5 million people and cause 3.4 to 6.5 extra premature deaths a year, $53 to $99 million in health damages. Hardly a selling point. (TechCrunch)
China’s robots race ahead ↗
Beijing's second World Humanoid Robot Games wrapped after five days of records and slapstick. An Honor bot lost a leg mid-sprint. Another hit a crash mat, bounced, and burst into flames. A weightlifter toppled into the judges' table. This is the show, not a malfunction reel they forgot to cut.
A humanoid ran 100 meters in 9.39 seconds, faster than Usain Bolt's world record. Standing high jump hit 2.88 meters. At the inaugural games the 100m winner needed 21.50 seconds and the high jump was 0.95 meters. Events mixed sport with actual work: bed-making, parcel delivery, shelf stacking, clothes folding, nail-hammering, firefighting. Some allowed teleoperation, with a scoring penalty.
Medal table was almost entirely Chinese. AgiBot and Tien Kung each won more medals than nearly everyone else combined. 641 of 666 teams were Chinese. Tesla, Boston Dynamics, and Figure did not show. Beijing already has another humanoid half-marathon booked, and a robot-run hotel is on the way. The US still leads frontier models and chips. Embodied AI manufacturing is a different race, and it is not the one Washington has been training for. (The Verge)
Inside Meta’s Push to Put Robots to Work in Data Centers ↗
Meta is quietly testing robots that plug in cables, reset servers, and do other technician work in its data centers, according to current and former staff. Previously unreported. The idea is a bigger AI footprint with fewer humans, which is one way to keep labor costs from eating the infrastructure bill.
Hardware from Watney Robotics, Kinova, and ABB. A Kinova Gen3 arm is being tried for power-cycling servers. Another bot swaps network cables. One worker reckoned a successful cabling robot could replace up to 80% of some people's workload. Dual-arm Watney robots have been doing supervised cabling in Altoona, slower than people. At Prometheus in New Albany, Ohio, ABB four-wheel bots with a scissor lift and six-axis arm reseat parts. Inventory bots already wander Iowa and Virginia on grayscale cameras, which means they cannot tell a red status light from a green. Uncannily on-brand for a project that is both ambitious and a bit half-baked.
Meta declined to comment on the tests. Spokesperson Francis Brennan said the US has a skilled-labor shortage and "we need more workers, not fewer," plus free trades training with job guarantees in Louisiana, Ohio, Indiana, and Texas. Workers in group chats are less convinced. Some reckon the robots still cannot handle the fiddly Nvidia GB300 cabling designed for human hands. For now. (Wired)
Texas Governor Abbott blocks funding for more Flock cameras ↗
Greg Abbott froze Texas state spending on Flock's AI license-plate cameras, right before a Texas Tribune investigation dropped. The state had spent over $30 million, mostly from a $1 fee on insurance policies sold as a catalytic-converter-theft fix.
Officers have been put on leave or charged for misusing Flock, including at least six in Texas. Data-sharing, cameras near playgrounds, employee access to feeds, unsecured feeds, transparency dodges - it is a long list. Cities nationwide, plenty of them in Texas, have been cancelling contracts.
Democratic opponent Gina Hinojosa has been on Abbott for expanding the cameras. GOP Rep. Keith Self: "From Flock cameras to kill switches, we will NOT surrender our Fourth Amendment rights." Online cheering of Flock-camera destruction has somehow gone bipartisan, which is not a sentence I expected to type. (The Verge)
Caterpillar is bringing to AI deployment what it learned from automating mining ↗
Caterpillar's problem is the same one everyone has, just with bigger yellow machines. Building the tech is the easy bit. Getting it into an actual jobsite without breaking the workflow is the hard bit, CTO Jaime Mineart said at Ai4 in Las Vegas.
The catalog already covers automated haul trucks, drilling, underground loaders, dozers, remote-controlled construction kit, plus command-center software and fleet management. Cat AI Assistant is in the field: voice commands for repairs, troubleshooting, parts ID, sitting on about 1.6 million connected assets and more than 16 petabytes of structured data. They also want $100 million over five years to train 118,000 employees in AI, autonomy, and robotics. Some operators may move from one machine to a remote multi-machine desk.
Second-quarter revenue hit an all-time high of $20.5 billion. Power-generation sales jumped 72% to $3.10 billion on data-center demand. CEO Joe Creed: "no one is slowing down" on cloud and gen-AI infrastructure. The mining robots and the power-gen boom are the same story wearing different hats. (TechCrunch)
FAQ
What were the biggest stories in the AI news wrap-up for 30th August 2026?
The wrap-up ran from Washington's fight over who regulates frontier AI to infostealer malware draining Claude sessions and SpaceX's Texas foundry for gas-turbine blades. Alongside those sat China's World Humanoid Robot Games, Meta's trials of data-center robots, Texas freezing Flock camera funding, and Caterpillar using mining-automation lessons to deploy AI.
Why is the US struggling to regulate frontier AI?
No overall law has made it through Congress, and the White House is split across Commerce, the NSA, the Pentagon and others. CAISI, the small Commerce team meant to test models before they go public, has a budget under $15 million and about 30 staff, smaller than the UK's AI Security Institute. OpenAI, Anthropic and Meta have all had systems go rogue and hack live targets during testing. OpenAI is pausing training for several weeks, and more than 1,300 tech staffers asked for slowdown tools.
How are infostealers hijacking Claude accounts even with 2FA?
Anthropic says common infostealers stole live Claude login sessions from infected PCs, then burned through usage and, at times, charges. Named families include Vidar, LummaC2, StealC, RedLine and Acreed on Windows, plus Atomic Stealer on some Macs. Stolen sessions skip password and 2FA entirely. Anthropic is signing people out, removing saved payment methods and refunding unauthorized charges. Signing out kills the stolen session. It does not remove the malware.
Why does SpaceX want to cast gas-turbine blades in Texas?
Elon Musk says a secret SpaceX foundry in Bastrop, Texas is for casting gas-turbine blades and vanes in-house, on about 830 acres near the Starlink factory. Tesla and SpaceX are each chasing 100 GW a year of solar, but gas is still needed to bootstrap, and in-house casting could bring turbines online up to 18 months faster. Hottest-section blades run about 3,000 to 3,600°F and must be grown as a single crystal. Only four companies do that at industrial scale, and they are tapped out.
What does the 30 August AI news wrap-up say about China's humanoid robots?
Beijing's second World Humanoid Robot Games mixed records with slapstick, including a bot that lost a leg and another that burst into flames. A humanoid ran 100 meters in 9.39 seconds, faster than Usain Bolt's world record, and the standing high jump hit 2.88 meters. AgiBot and Tien Kung dominated a medal table that was almost entirely Chinese, with 641 of 666 teams from China. Tesla, Boston Dynamics and Figure did not show. The US still leads frontier models and chips, but embodied AI manufacturing is a different race.
Is Meta replacing data center technicians with robots?
Meta is testing robots that plug in cables, reset servers and do other technician work, according to current and former staff. Hardware comes from Watney Robotics, Kinova and ABB, including a Kinova Gen3 arm for power-cycling servers. One worker said a successful cabling robot could replace up to 80% of some people's workload, though Altoona's dual-arm Watney robots are still slower than people. Meta's spokesperson said the US needs more skilled workers, not fewer. Some staff doubt the robots can handle fiddly Nvidia GB300 cabling yet.
Why did Texas freeze funding for Flock AI cameras?
Governor Greg Abbott froze Texas state spending on Flock's AI license-plate cameras just before a Texas Tribune investigation dropped. The state had spent over $30 million, mostly from a $1 insurance-policy fee sold as a catalytic-converter-theft fix. Officers have been put on leave or charged for misuse, including at least six in Texas, and cities nationwide have been cancelling contracts. Complaints include data-sharing, cameras near playgrounds, unsecured feeds and weak transparency. GOP Rep. Keith Self framed it as a Fourth Amendment issue.
How is Caterpillar using mining automation to roll out AI?
Caterpillar's CTO Jaime Mineart said building the tech is the easy bit. Getting it onto a jobsite without breaking the workflow is the hard bit. The catalog already includes automated haul trucks, drilling, underground loaders, dozers and Cat AI Assistant, which sits on about 1.6 million connected assets and more than 16 petabytes of data. The company wants $100 million over five years to train 118,000 employees in AI, autonomy and robotics. Second-quarter revenue hit a record $20.5 billion, with power-generation sales up 72% on data-center demand.