Newsletter

Fugaku stays offline as RIKEN finds stolen SSH keys and 4,600 exposed accounts

Japan's flagship supercomputer was breached through a public status dashboard, and the intruders left with credentials that reached other organizations' systems.

Rows of tall black computer racks with blue front panels lettered Fugaku and Fujitsu in a bright white machine room, behind a low chain barrier.
Lead image Fugaku's racks at the RIKEN Center for Computational Science in Kobe. The machine has been shut down since November 14. · Photo: Barsaka2 / Wikimedia Commons · CC0 1.0
On this page

Fugaku, Japan’s flagship supercomputer, has been offline since the early hours of November 14. RIKEN cut external access after confirming that an outside party had reached the network that runs the machine. The intruders entered through a public web page showing Fugaku’s operating status, then moved into the internal network behind it. They tampered with several servers and installed malicious programs.

On Friday, RIKEN published a second notice describing what was taken. Investigators found traces that an intruder reached registration records for 4,222 Fugaku users and 375 users of RIKYU, a smaller system for AI research. Names, email addresses, user IDs and contact details were exposed. The intruders also removed 430 authentication credentials from user home directories. Of those, 233 were SSH private keys stored without a passphrase.

RIKEN says it has found no confirmed misuse, but it is treating the data as breached. Some stolen keys were used in attempts to reach systems outside RIKEN, and affected users have been asked to replace them. Fugaku will restart only after public services are separated from the administrative network and the servers are rebuilt. RIKEN has notified police in Saitama and Hyogo prefectures.

Featured source: RIKEN second notice , RIKEN first notice .

A canary experiment finds leaks across the AI inference chain

HD Moore, chief research officer at the security firm Dragos, spent five months sending about two million prompts through commercial AI services. Each carried a secret that existed nowhere else, such as a web address or an AWS key tied to an account he controlled. Anything that later used a secret showed where the prompt had travelled.

The results describe three problems. Data from at least three hosting providers was exposed, and the stolen test keys were used in tens of thousands of AWS API calls. Several marketplace resellers logged customer prompts without exploiting anything, and some providers appear to have ignored no-retention flags. A Chinese lab fetched secret URLs from more than half the prompts sent to its models.

Filed from: Dragos , Ars Technica .

Ruby fixes a pre-authentication path to code execution in default Rails apps

Ruby released versions 3.3.14, 3.4.13 and 4.0.10 on November 18 to fix seven memory-safety flaws in its regular expression engine and in string case conversion. The Ruby team said it knew of proofs of concept that achieve remote code execution.

The most serious flaw is a one-byte heap overflow triggered when certain strings are converted to uppercase. Cloudflare, which calls the bug ruby2shell and lists it as CVE-2026-132825, says Rack’s method-override middleware uppercases a request parameter before any authentication check. That puts default Rails and classic Sinatra applications at risk, and Ruby 3.2 and earlier will not be fixed.

Filed from: Cloudflare , Ruby advisory .

SpaceX lenders try to keep their chip collateral on Earth

SpaceX sold more than $30 billion of a $43 billion debt package on Thursday, the Financial Times reported. The largest part is a $25 billion investment-grade bond secured by five-year lease payments for Nvidia chips. The debt sits in a special purpose vehicle that keeps it off SpaceX’s balance sheet.

Lenders raised an unusual concern: Musk wants to put data centers in orbit, which could mean launching the very chips that secure their loans. People familiar with the talks said bondholders tried to bar that, and it is unclear whether the final documents include the restriction. Moody’s rated the debt A3, one notch above SpaceX itself.

Filed from: Ars Technica .

Meta reopens its 2030 net-zero goal as data-center emissions climb

Meta’s 2026 sustainability report says the company is re-evaluating the “timelines, sequencing, and interim milestones” of its climate goals, including the 2030 net-zero target. Executives attributed the rethink to demand for AI infrastructure growing faster than almost anyone projected.

The report put Meta’s emissions at more than 21.7 million metric tons of carbon dioxide equivalent in 2025, up from 15.6 million the year before. That is a rise of nearly 40 percent, the steepest among the four largest US technology companies. Microsoft reported 25 percent, Google 18 percent and Amazon 16 percent.

Filed from: Ars Technica , Meta sustainability report .

Google signs a 50 MW superhot geothermal deal for Oregon data centers

Quaise Energy said Google signed a power purchase agreement for 50 megawatts from Project Obsidian, its superhot geothermal project in central Oregon. Power is due as early as 2030 and would serve Google’s data center operations in The Dalles.

The first phase will use conventional drilling to reach superhot rock at commercial scale. The second is meant to be the first worldwide to drill with millimeter waves, which Quaise says can reach such heat almost anywhere. TechCrunch reported that Google agreed to pay more than $500 million.

Filed from: Quaise Energy , TechCrunch .

METR finds AI time horizons still doubling as judgment trails behind

METR’s capability update adds harder, more open-ended tasks to the measurement behind its widely shared time-horizon charts. On the combined task sets, the length of work a model can finish with 50 percent reliability doubles roughly every 3.6 months.

The new tasks show where models still stumble. Asked to design evaluations that METR researchers would find useful, no model produced one that qualified. Models also scored below expert humans when judging philosophical arguments, and model judges overestimated work by strong models. METR says public evidence cannot yet rule out drastic acceleration of AI research.

Filed from: METR , MIT Technology Review .

From the Community

Paper and Lean formalization claim finite-time blow-up for Navier–Stokes

A repository from L. Alpöge and T. Buckmaster hosts a 79-page paper, dated November 20, that claims finite-time blow-up for the Navier–Stokes equations. It also holds a Lean 4 formalization built on Mathlib.

The repository says the formal proofs cover the paper’s main theorems. It also states two omissions. The numerical certificate named in the paper is not included, and the interval-arithmetic check of the paper’s conditions sits outside the formalization. Both are promised later, so the claim rests on material that readers cannot yet inspect.

Filed from: Paper and Lean formalization repository .

Thinking Machines releases a laptop-sized open-weights model

Thinking Machines Lab released Inkling 2 Mini, a mixture-of-experts model with 26.4 billion total parameters and 3.4 billion active. It runs offline on a consumer laptop in about 20GB of memory. Its middle layers run three times for each token, a looping design meant to buy more reasoning without a longer written chain of thought.

The company reports that it leads open-weights models under 100 billion parameters on agentic coding tests, with 48.2 percent on DeepSWE v1.1. Those figures come from its own release, and evaluation setups vary between models.

Filed from: Thinking Machines Lab .

Continue reading

Complete index →