Amazon-Backed Texas Gas Plant Receives Permit for Up to 33 Million Tons of CO2

Amazon has invested in the construction of a natural-gas power plant in Pecos County, Texas that would primarily power a new data center and, under the state permit it received, could become one of the largest single sources of power-plant carbon pollution in the United States. The plant’s 35 gas turbines would generate 7.65 gigawatts of electricity, initially without connecting to the Texas grid, and the permit allows emissions of up to 33 million tons of CO2 — more than even the largest coal plant in the country, according to Cleanview, which tracks data centers and their associated power projects.

Amazon confirmed that it purchased the site and plans to buy power from GW Ranch, the project’s name. Plants rarely emit as much as their permits allow, The Verge notes, and Amazon spokeswoman Margaret Callahan told the New York Times that while “the world looks different now than when we co-founded the climate pledge,” the company’s commitment to it has not changed. Amazon is one of several hyperscalers — including Meta and Google — building their own gas-fired generation as AI data center demand climbs, a trend the Trump administration has sought to ease by lifting restrictions on such plants.

Featured source: The Verge , The New York Times .

Other Stories

GPTZero Investigation Finds Hallucinated Citations in Four PwC Middle East Reports

GPTZero’s investigation team says four thought-leadership reports published by PwC Middle East between 2024 and 2026 contain hallucinated citations, fabricated claims, and AI-generated or AI-assisted text. The findings include a 2025 report promoting a framework called “Citizen Pulse” that the investigators say appears not to exist outside the report, yet claims governments in Denmark, Saudi Arabia, the United States, and Australia use it; none of the cited sources support the claim. PwC told the Financial Times that it has quality-control processes for research and content development that all staff are expected to follow. GPTZero says the past 12 months of scanning documented similar citation issues at all four big accountancy firms, and the investigators acknowledge that their detection tools cannot classify every imperfect citation as a hallucination.

Filed from: TechRadar , GPTZero investigation .

Adversarial Patterns Defeat Surveillance Camera Detection in First Public Test

Security researcher Bill Swearingen has built a reinforcement-learning system that generates printed patterns which block detection by some of the most widely deployed surveillance cameras and license plate readers. After roughly 31 million tests, he says the patterns defeat all 11 open-source detection algorithms he tested, including software powering Flock license plate readers, Axon body-worn cameras, and cameras running Clearview AI. In a first public test on Friday at the Def Con conference in Las Vegas, a 2009 Toyota Yaris covered in one of the patterns avoided detection by a Flock camera, though Swearingen said the wheels remained a challenge. The patterns scramble object, face, and vehicle detection without stopping cameras from recording; Swearingen, who co-founded the Kansas City cybersecurity meet-up SecKC, says he is keeping his strongest patterns offline so camera makers cannot defeat them.

Filed from: TechCrunch , noRecognition project .

Nvidia Open-Sources cuFile and Launches Storage-Next With More Than 40 Vendors

Nvidia has open-sourced its cuFile APIs and the storage stack beneath them into a new GitHub organization with Google, Intel, and Meta as founding maintainers, and formally launched Storage-Next with more than 40 flash and storage vendors. The company also named SCADA, a framework that moves the storage control path onto the GPU so parallel GPUs can construct and complete their own storage requests, absorbing per-operation latency. The work targets what Nvidia describes as a 512-byte problem: AI inference fetches embeddings and KV cache blocks in a few hundred bytes, while enterprise SSD controllers tuned for 4KB reads impose roughly eightfold read amplification at that size. cuFile has shipped since 2021 as part of GPUDirect Storage; the new stack keeps the CPU out of the data path for bulk transfers while SCADA handles high volumes of small random reads.

Filed from: TechRadar .

Ransomware Gangs Target Mid-Level Managers Rather Than Executives

Zscaler’s ThreatLabz researchers tracked 351 victims across 334 organizations caught up in a single ransomware campaign over the course of a month and found attackers increasingly targeting mid-level managers instead of the C-suite. Nearly two-thirds of victims held manager-level titles or above, the average victim was a 46-year-old Gen Xer, three-quarters worked in accounting and finance, sales, operations, HR, or marketing, and half worked in the industrial or IT sectors. Zscaler says attackers map reporting lines from compromised systems and public data to reach employees with the authority to accelerate payment decisions — “business privilege” rather than technical privilege, since managers approve invoices, budgets, and contracts. The report also says ransomware attempts blocked across Zscaler’s cloud platform increased 146 percent over the past year, public extortion cases rose 70 percent, and the volume of data stolen from victims climbed 92 percent.

Filed from: The Register , Zscaler ThreatLabz .

Claude Code’s Auto Mode Becomes the Default on Pro, Max, and Team Plans

Anthropic is making auto mode the default in Claude Code, starting August 14 for new sessions on Pro, Max, and Team plans, enabling longer-running autonomous work while routing each tool call through a classifier that blocks irreversible, destructive, or out-of-environment actions. Users who set a different default may get a one-time prompt about switching, and pinned defaults are unchanged. The classifier costs a small number of extra tokens per tool call, and Anthropic says it is no longer charging users on those plans for that overhead, effective immediately. Auto mode remains opt-in for now on Claude Enterprise, the API, and cloud platforms including Amazon Bedrock and Google Cloud’s Agent Platform, with plans to make it the default across all of them in the coming month.

Filed from: Anthropic .

London’s King’s Cross Has Become One of the World’s Leading AI Hubs

King’s Cross, a London district redeveloped from a rundown area over the past two decades, now hosts OpenAI, Meta, Isomorphic Labs, and other AI companies and is widely described as the U.K.’s AI heartland. Dealroom counts roughly 3,600 AI startups in London, which have raised about $12.1 billion of the $14.8 billion raised in the city since late July. Real estate firm Knight Frank says AI-related startups have leased more than 1 million square feet of office space in London since the start of June, and the vacancy rate for conventional office space in the area is 0.9 percent. The district’s draw dates to 2016, when Google-acquired DeepMind moved in; some founders and investors frame the cluster as a response to European AI dependence, citing Anthropic’s shutdown of access to its Mythos and Fable models this summer as a reminder that Europe needs to build and hold some of its own AI capacity.

Filed from: TechCrunch .

Israel Completes First Test of a Wildfire-Detection Drone

Israel has completed the first test of a drone designed to detect wildfires within minutes of ignition and alert the country’s Fire and Rescue Authority with automatic location data. Built around Israel Aerospace Industries’ APUS 25 platform, the 24-kilogram quadcopter pairs onboard cameras and infrared sensors with software trained to spot heat signatures even through smoke or darkness, and draws power from a combustion engine rather than a battery, giving it up to five hours of endurance. The project brings together Israel’s Ministry of National Security, the Innovation Authority, IAI, and the Technion. Further testing is planned across a broader mix of weather and terrain before any wider rollout, per the partners’ announcement.

Filed from: TechRadar .

From the Community

CSS Attacks Against Webmail Clients

PortSwigger researcher Gareth Heyes details novel CSS-based attacks against webmail clients such as Gmail, Outlook, and ProtonMail, which render untrusted CSS inside a trusted UI. Techniques include abusing HTML labels to control the interface, exfiltrating tokens via attribute selectors, and bypassing content security policies to steal passwords; the research demonstrates real-world impact, including account takeover on Medium and controlling OpenAI’s Atlas browser via email. The paper highlights how hard it is to sanitize untrusted CSS in a trusted context.

Filed from: PortSwigger Research .

Reviving a Four-Year-Old reMarkable 2

Developer Oskari Mantere documents reviving a reMarkable 2 tablet that had sat unused for four years, with software updates and cloud sync both failing. The fix involved SSH access, correcting the system clock so updates would download, then manually restarting the update services to jump from an obsolete version to the latest release; the post also notes that newer software versions silently disable SSH over Wi-Fi, requiring USB-based access and a command to re-enable it.

Filed from: Oskari Mantere .

Microsoft Word 1.1a Ported to Native Windows x64

A developer has produced a native Windows x64 port of Microsoft Word for Windows 1.1a, building the original source and resources together with modern replacements for its 16-bit assembly, segmented-memory, and Win16 platform boundaries. The repository describes the result as the original application running as a 64-bit Windows executable rather than an emulator or a reimplementation, and documents the build and its extra research features.

Filed from: GitHub repository .

Multistack Concatenative Programming Languages

An essay on concatenative programming languages — the style centered on point-free composition where “juxtaposition is composition” — surveys how the family has evolved beyond the single data stack, examining languages that add auxiliary stacks, lexical variables, and other ergonomics to tame the “write-only” code that pure stack programming can produce. It works through the trade-offs of each design as a tour of a largely unexplored corner of language design.

Filed from: sheeeeeeeep.art .