Newsletter

Frontier AI Agents Stall on Open-Ended Research, Princeton-Led Study Finds

In shadow evaluations, frontier agents completed the engineering of two unpublished NeurIPS papers but failed their research questions, while Anthro Energy broke ground on a Kentucky factory aimed at solid-state battery electrolytes.

A close view of server racks in a research-institution computer room, dark chassis with glowing blue status lights receding into soft focus.
Lead image Server racks at the NOIRLab headquarters computer room in Tucson, Arizona, illustrative of the large-scale compute used in frontier AI research evaluations. Image: NOIRLab/NSF/AURA/T. Slovinský, CC BY 4.0. · Image: NOIRLab/NSF/AURA/T. Slovinský · CC BY 4.0
On this page

A multi-institution research team led by Peter Kirgis and Sayash Kapoor at Princeton tested whether frontier AI agents can conduct open-ended AI research — the kind of free-form investigation, requiring judgment and taste, that underpins forecasts of AI improving itself. In “shadow evaluations,” the team gave agents the central research question of two high-quality unpublished NeurIPS 2026 submissions, plus six days and thousands of dollars of compute, and had the papers’ original authors grade the results. The agents completed all of the engineering without human help but made no substantial progress toward the research questions, and both papers were rejected by their authors.

The researchers identify five recurring failure modes, including poor judgment about the bar for publishable research, uncreative responses to shortcomings in research design, and ineffective backtracking from dead ends. The finding suggests that some hyped timelines for automating AI research — which depend on models that can write code, generate training data, and optimize their own chips — are running ahead of the evidence, though the study is based on two case studies rather than a broad evaluation.

Featured source: MIT Technology Review , arXiv paper .

Tracking Device Confirms Amazon Buys Rare Books, Scans Them for AI Training, and Destroys Them

404 Media placed a tracking device in a rare book it suspected would be bought for AI training data and followed the shipment to an Amazon warehouse in Las Vegas, where employees receive bulk shipments of printed books, cut off the bindings, and scan the pages for training data, destroying the books in the process. Amazon declined to comment on AI training specifically, saying only that it “purchases books through commercial channels to help develop and improve the products and services our customers use.” Rivals Anthropic and xAI have publicly stated they are not training on rare or antique books.

Filed from: Ars Technica , 404 Media investigation .

Anthropic’s Annualized Revenue Run Rate Passes $65 Billion

Bloomberg reports that Anthropic’s annualized revenue run rate surpassed $65 billion at the end of July, up from $47 billion in May and about $9 billion at the end of last year, and the Financial Times reports investors expect 2026 revenue between $100 billion and $120 billion. Anthropic has filed confidential IPO paperwork and could list as soon as this fall, per the FT; it was last valued at $965 billion in late May, when it raised a $65 billion round. Bloomberg separately reported that OpenAI's run rate doubled to $40 billion from $20 billion at the end of 2025, and TechCrunch notes the two companies may calculate revenue metrics differently.

Filed from: TechCrunch , Bloomberg Technology .

Nvidia Discloses a Roughly $21 Billion Stake in SpaceX

Nvidia’s Friday SEC filing shows it owned nearly 123 million SpaceX shares worth about $21 billion at the end of June; because SpaceX shares have fallen since the company's June initial public offering, Ars Technica notes the stake would be worth about $17 billion now. The holding grew out of Nvidia’s January investment in xAI, completed shortly before Elon Musk combined the AI lab with SpaceX. At SpaceX’s first public earnings call, Musk said the company builds exclusively on Nvidia, calling its Vera Rubin architecture “the best AI computer.”

Filed from: Ars Technica , SEC EDGAR — Nvidia filings .

Nvidia to Invest $1.5 Billion in the Data Center Developer Behind OpenAI’s Ports-Pike Project

Nvidia said it will invest $1.5 billion in SB Energy, the SoftBank- and OpenAI-backed data center developer, in a deal that makes Nvidia the sole supplier of compute infrastructure at OpenAI's Ports-Pike campus near Cincinnati, Ohio, with up to $105 billion in credit to help build the facility. The project is planned to scale from an initial 4.25 gigawatts toward 8 gigawatts, per documents Nvidia filed with the SEC, and includes a 9.2-gigawatt natural-gas power plant on Department of Energy land that previously enriched uranium. BloombergNEF pegs the plant alone at an expected $33 billion, reflecting a 66% rise in gas-plant construction costs over two years.

Filed from: TechCrunch .

Anthro Energy Breaks Ground on a Kentucky Factory for Battery Electrolytes, Including Solid-State

Anthro Energy broke ground on a factory in Louisville, Kentucky, with 25 gigawatt-hours of annual electrolyte capacity, enough battery materials for more than 300,000 electric vehicles, with production scheduled to start in 2028. CEO David Mackanic says the plant will serve U.S. customers seeking electrolyte supply not controlled by Chinese companies — “FEOC-free,” in industry terms — noting that 70% of U.S. battery production facilities sit within a 12-hour drive. The project draws a $24.9 million Department of Energy award under the Bipartisan Infrastructure Law, $18.4 million in Inflation Reduction Act investment tax credits, and $2.3 million in Kentucky incentives tied to 110 permanent jobs, and is designed to eventually emphasize the company’s Proteus polymer electrolyte for solid- and semi-solid-state batteries.

Filed from: TechCrunch , Anthro Energy news .

Groq Raises $350 Million at a Lower Valuation as It Pivots to Nvidia-Powered Clouds

Groq raised $350 million, led by investment firm Disruptive with planned participation from Nvidia, at a $3.5 billion valuation — down from the $6.9 billion it was worth last September, before Nvidia hired founder and CEO Jonathan Ross under a $20 billion licensing deal. The company, which pivoted from its own inference chips to operating Nvidia-powered clouds, says it runs 13 data centers across four regions serving more than 6 million developers, and plans to scale from 54 megawatts to over 200 megawatts in 2027. Groq’s spokesperson framed the lower figure as a fresh valuation for the post-licensing company rather than a down round.

Filed from: TechCrunch .

GitHub Outage Hits a 50% Error Rate on Repository Downloads

GitHub’s status page reported a 50% error rate on repository content downloads beginning 13:40 UTC Monday, with degraded availability across Issues — which was itself experiencing an outage — and GitHub Copilot; the cause was unknown at the time and the company was investigating. The Register notes the incident continues a pattern of disruptions: Actions and Pages failed earlier in August, and GitHub’s May availability report acknowledged strain from AI-assisted coding and agentic workflows. In June, software engineering SVP Jakub Oleksy promised structural changes to “permanently remove failure modes.”

Filed from: The Register , GitHub status .

“Unprecedented” Number of Apple Users Received Spyware Alerts, Investigators Say

Apple sent a new wave of threat notifications Friday warning customers in 110 countries they had been targeted with suspected mercenary spyware, and digital-rights group Access Now says it has received a record number of reports to its helpline — around 30–40% more than it typically sees after Apple’s past alert batches. Citizen Lab senior researcher John Scott-Railton called the scale and geographic diversity of public reports “pretty unprecedented.” Experts caution the increase partly reflects Apple’s broader alerting methods introduced this year — notifications now appear on the lock screen, in Settings, by email, and at web login — rather than purely a rise in attacks.

Filed from: TechCrunch , Apple: about mercenary spyware attacks and threat notifications .

Germany’s Federal Cartel Office issued a preliminary assessment finding that Apple’s App Tracking Transparency prompt is designed to steer users away from consenting to third-party data use, while the Personalized Ads prompt steers them toward consenting to Apple’s advertising, citing the warning-hand icon, differing wording, ambiguous data-use descriptions, and the order of selection options. The regulator, acting under the EU’s Digital Markets Act, gave Apple four months to update the design and language; per Reuters, the changes would apply in almost all European Union countries. The proposed design would make third-party prompts neutral and let developers combine or link Apple’s pop-ups with other consent prompts.

Filed from: The Verge , Bundeskartellamt .

Pokémon Center Notifies UK and Germany Customers of a Breach at Logistics Partner CEVA

Pokémon Center is notifying customers in the United Kingdom and Germany that hackers stole personal and order information from CEVA Logistics, the third-party provider that ships PokemonCenter.com orders; the company says unauthorized parties may have obtained full names, mailing addresses, phone numbers, email addresses, and details of order contents, and that CEVA does not hold payment card details. Pokémon Center canceled some UK orders and is warning of shipping delays. CEVA, a subsidiary of shipping group CMA CGM, says attackers breached its systems between July 29 and August 1, disrupting eight of its European warehouses and affecting multiple retailers, including Valve, which told European Steam hardware customers their data was stolen; no group has claimed responsibility.

Filed from: BleepingComputer , TechRadar .

Siemens and Reinhausen Develop an 800 VDC Solid-State Transformer for AI Data Centers

Siemens and fellow German engineering firm Reinhausen are developing a modular solid-state transformer that converts medium-voltage AC at up to 36 kV into a stable 800-volt DC output for high-density AI racks, a design the companies say is compact and galvanically isolated, cutting the intermediate conversion stages of conventional power delivery; they have not said when it will be commercially available. The Register notes 800 VDC distribution is the direction Nvidia and others favor for racks drawing a megawatt or more, with Nvidia’s MGX-compatible 800 VDC rack expected in the second half of 2026, and Infineon positioning silicon-carbide components for the standard.

Filed from: The Register , Siemens press newsroom .

AI Models Get Convenient Amnesia About Source Material as They Grow, MIT Researchers Find

MIT CSAIL researchers Zheng Dai and David K. Gifford report in Nature Communications that attributing a diffusion model’s output to specific training data becomes harder, not easier, as models scale — a phenomenon they call attribution decay: in ablation tests on very large models, removing an image such as the Mona Lisa or all of Leonardo da Vinci’s work still left the model able to reproduce the image or style. The finding complicates copyright disputes such as Andersen v. Stability AI, where plaintiffs seek training datasets, and could cut both ways: Gifford argues it raises questions about fair use and output copyrightability, while the paper also suggests scaling up is a potential liability-avoidance strategy.

Filed from: The Register , Paper (Nature Communications) .

A new “KDE Software Initiative” promises at least three years of backported fixes for KDE Plasma 6.6 and the associated KDE Frameworks 6.24 and KDE Gear 25.12, effectively making the desktop, which debuted six months ago, a long-term-support release. The effort targets Kubuntu 26.04, the latest LTS release of Ubuntu’s KDE flavor; hardware vendor Kubuntu Focus is funding KDE developer Nate Graham’s company TechPaladin to do the maintenance, and other distributions are free to adopt the maintained components. The Register notes the arrangement matters because only Ubuntu’s official GNOME edition gets the full five-year support lifespan, while Kubuntu, like the other flavors, receives three years of desktop support.

Filed from: The Register , Nate Graham’s blog .

From the Community

Wiz Red Agent Exploits a Snowflake Vulnerability Missed by GitHub Copilot Autofix

Wiz Research’s autonomous AI security agent, Red Agent, discovered and exploited a critical GitHub Actions script injection vulnerability in Snowflake’s public repository, gaining read access to Snowflake’s internal Jira. The vulnerability was introduced by a pull request co-authored by Copilot Autofix, which replaced a safe input-handling pattern with direct string interpolation, and GitHub’s AI review failed to flag it. Snowflake patched the vulnerability the same day, rotated the affected credential, and confirmed via audit logs that Wiz was the sole actor during the five-day exposure window.

Filed from: Wiz Blog .

CSS: The Bomb Inside Your Inbox

PortSwigger researcher Gareth Heyes published a detailed study of attacks against webmail clients built on sanitized CSS, demonstrating how it can be abused to control UI, exfiltrate tokens, compromise third-party websites, and even steal passwords. The research covers label-based manipulation of Outlook’s interface, indirect prompt injection in OpenAI’s Atlas browser via email, and a clipboard-based attack that steals Medium login tokens when pasted into a draft, plus token exfiltration even when CSP blocks external resources.

Filed from: PortSwigger Research .

Israel Creates a Fake Think Tank to Influence AI Chatbots

Responsible Statecraft reports that the Hanover Institute for Public Policy, a fake think tank created on behalf of the Israeli Government Advertising Agency, published more than 100 articles in just over a week, engineered to influence AI chatbots such as ChatGPT and Gemini. The articles mimic credible think-tank reports with footnotes and neutral tone and are produced by Piro, Inc., which received $900,000 from the Israeli government; NewsGuard analyst Alice Lee noted the site’s perfect mimicry of a typical American think tank, and GPTZero flagged 11 of 12 analyzed articles as AI-written with high confidence.

Filed from: Responsible Statecraft .

GPU Offload in Rust: Portable, Safe, and Fast

A new arXiv preprint presents a zero-overhead, multi-vendor GPU compilation framework built natively into the Rust compiler and LLVM backends, leveraging Rust’s ownership and strict aliasing guarantees to manage data transfers through LLVM’s Offload infrastructure and addressing cross-vendor ABI lowering mismatches with a two-pass compilation pipeline. Evaluation on the RAJAPerf suite shows competitive kernel performance against native, hand-optimized CUDA and HIP C++ baselines.

Filed from: arXiv .

Continue reading

Complete index →