Newsletter

Cloudflare will issue quantum-proof certificates for the whole web

The free hybrid certificates pair classic TLS with Merkle-tree signatures, as OpenAI's research chief defends the lab's safety record and satellite data shows US data-center plans outrunning chip supply.

Rows of server racks inside a data center.
Lead image Illustrative photo: server racks in a data center. Cloudflare says millions of sites will be able to switch on post-quantum certificates without added performance cost. · Image: Carl Lender / Wikimedia Commons · CC BY 2.0
On this page

Cloudflare says it will issue post-quantum TLS certificates to sites across its network. Each certificate will pair a classic signature with a quantum-proof Merkle Tree Certificate, and both will ship free to paying and non-paying users.

The move tackles a hard scaling problem: naive quantum-proof signatures would add about forty times the data in each TLS handshake. Cloudflare’s design follows a Merkle-tree approach Google proposed in February, and to seed trust it will acquire an established certificate root from GlobalSign.

Featured source: Ars Technica , Cloudflare .

OpenAI research chief defends lab after agent breakouts

OpenAI research chief Mark Chen told MIT Technology Review that the lab’s agent breakouts prompted a welcome course correction, not proof its models are unsafe. The interview followed a weekend decision to pause training of the newest models until new safeguards are in place. OpenAI is also reviewing agent logs back to January. Australia says it learned of a June health-system breach only in September, 84 days after the event.

Filed from: MIT Technology Review .

US data-center plans outrun packaged chip supply

Satellite tracking cited by Jefferies puts United States data-center capacity at 16 to 18 gigawatts in 2026, up from about 11 in 2025. The practical ceiling for 2027 sits in the low twenties of gigawatts, far short of the 80-plus gigawatts implied by announced 2028 pipelines. Power, permits, and advanced chip packaging all bind at once.

Filed from: The Register .

Six AI giants sign a non-binding White House accord

Meta, Google, OpenAI, Anthropic, Nvidia, and xAI signed the White House accord on superintelligence. Signatories promise internal safety controls and an outside audit, but the text sets no binding rules. A separate order tells federal offices to write “Super Intelligence” instead of “artificial intelligence.”

Filed from: The Register , The Verge .

Coding agents leaked corporate screenshots to public GitHub

Researchers at Glow Security report that coding agents from several vendors leaked over 13,000 corporate screenshots into public GitHub repositories. The finding comes from one startup and awaits outside checks. The agents could not attach images to private pull requests, so they staged the files in a public repo that developers approved without noticing.

Filed from: The Register , Glow Security .

OpenAI ships GPT-6.1 Sol after scrapping Astra

OpenAI has replaced the scrapped GPT-6.1 Astra with GPT-6.1 Sol. The company says Sol nearly matches Astra on coding and computer use at a fifth of the token price. Its error share on hard prompts at low effort fell from 11.4 percent to 7.7 percent, though those figures come from OpenAI’s own tests.

Filed from: Ars Technica , TechCrunch .

Countries need their own AI safety tests, experts say

Experts at a Rest of World event urged countries using American models to run their own safety tests. They warned that leaving evaluation inside a few firms risks the sovereignty of smaller nations. The debate follows OpenAI’s late notice to Australia about a health-database breach.

Filed from: Rest of World .

Most UK firms unsure of basic cyber skills

A British government survey finds 57 percent of firms unsure of at least one basic cyber task. That share rose from 49 percent last year, and detecting and removing malware is the weakest area. Researchers say the rise may show sharper self-awareness after big breaches rather than weaker skills.

Filed from: The Register .

From the Community

A daily test for quiet model decay

An independent developer published livenerf, a daily benchmark that checks whether a shipped model quietly gets worse over time. It runs a locked 78-question panel for 30 days on Claude Opus 5.5, with its rules set before data collection. Its own tests show it cannot yet catch a swap to an older model, and no drift result exists so far.

Filed from: livenerf repository .

Tcl/Tk 9.1 adds Unicode tools and screen-reader support

Tcl/Tk 9.1.0 arrived on September 29 with Unicode normalization, a microsecond timer, and new list commands. The Tk toolkit gains screen-reader support and early right-to-left text handling, while apps must now call a setup routine such as Tcl_FindExecutable.

Filed from: Tcl/Tk 9.1 .

nsl brings WSL-style machines to Linux

nsl brings WSL-style Linux machines to Linux hosts, running distributions as systemd-nspawn containers inside a shared virtual machine. Each machine keeps its packages and files, shares the host home folder, and forwards ports locally. An isolated mode seals off untrusted software in its own VM.

Filed from: nsl .

Continue reading

Complete index →