AWS declares its Bahrain cloud region a permanent wartime loss

Amazon Web Services says it cannot restore access to resources and data hosted exclusively in its Bahrain Region after wartime damage spanned multiple Availability Zones, exceeding what its regional and multi-AZ designs are built to withstand. After a first Bahrain zone was damaged in March, AWS advised customers to migrate elsewhere, and most did so before further attacks in April left the whole region unavailable. The company reached the same verdict for one of three zones in its UAE Region, while recovery work continues in the other two and most affected customers have restarted elsewhere from backups.

This appears to be the first permanent loss of a major cloud region to warfare, and it confirms the warning AWS issued after the first strikes: multi-AZ redundancy cannot survive a region-wide event, and only remote backups in another region save workloads. The damage is characterized by The Register, reporting on an AWS Health Dashboard update, as Iranian strikes including March drone hits on two UAE facilities amid retaliation for US-Israeli strikes on Iran; AWS declined to comment beyond the dashboard notice.

Featured source: The Register , AWS Health Dashboard .

Other Stories

Apple ships record patch cycle covering more than 260 CVEs

Apple has addressed more than 260 CVEs across its operating systems, browsers, and other software, the largest single patch cycle in the company’s history, with iOS 27 fixing a record 122 flaws and macOS 27 Golden Gate a record 204. Only about ten of the CVEs credit AI-assisted discovery, including two iPhone flaws found with Anthropic’s Claude, while the rest reflect conventional research. None of the vulnerabilities are listed as under active exploitation, though that status can change quickly once details are public.

Filed from: The Register , Apple support .

Cisco email gateways can be rooted by a single email

Attackers are actively exploiting a critical flaw in Cisco Secure Email Gateway appliances, tracked as CVE-2026-76461 with a 9.8 CVSS score, that lets an unauthenticated attacker run commands as root by sending a booby-trapped message through a vulnerable gateway. Cisco says there are no workarounds, only patching, and warns that intruders with root access may tamper with logs to cover their tracks, so a clean log does not prove a clean box. The company found the bug while resolving a support case, says it became aware of active exploitation in September, and has not attributed the attacks or said how many organizations are affected.

Filed from: The Register , Cisco advisory .

OpenAI, Anthropic, and Google have discussed AI safety for weeks

OpenAI policy chief Chris Lehane told reporters the company has been working with Anthropic and Google DeepMind on AI safety for several weeks, following Anthropic CEO Dario Amodei’s weekend essay calling for the industry to pace frontier development. The talks reportedly include a possible industry standards body, an idea Google’s Demis Hassabis proposed for July, and Lehane said OpenAI backs a FRONTIER Act provision requiring frontier labs to admit independent verification organizations. Participants acknowledge antitrust risk in coordination among competitors, and Lehane reportedly maintains no government waiver is needed, against a White House that dismisses slowdown calls as a hoax.

Filed from: TechCrunch , Bloomberg Law .

Spain records its first data breach caused by an autonomous AI agent

Spain’s data protection agency says an individual deployed an AI agent built on a known but unnamed large language model that scanned an organization’s files, broke into its systems, ran vulnerability scans, and gained read and write access to files holding personal data and invoices. Agency president Francisco Pérez Bes called for an immediate review of security and data protection models, arguing human supervision must be backed by detection and response fast enough to match agentic attacks. The model, the target, and other technical details remain undisclosed, so the account cannot yet be independently verified.

Filed from: The Register .

Poll finds 61 percent of likely voters oppose nearby AI data centers

A New York Times and Siena University poll of 1,503 likely voters in early September found 61 percent oppose building AI data centers, with only 14 percent strongly in favor, confirming a run of polls showing deep hostility to local AI infrastructure. Opposition crossed party lines, though Trump voters split nearly evenly while Harris voters and nonvoters leaned heavily against, and the most cited reason was environment and water use at 32 percent. Among opponents, 56 percent favored limits while 38 percent preferred a total ban.

Filed from: The Verge , Gallup .

US data centers may burn more gas than Germany and Japan combined by 2035

US data centers are projected to consume about 18 billion cubic feet of natural gas per day by 2035, more than Germany and Japan combined, according to a new BloombergNEF report, nearly double what the firm predicted nine months ago. Grid-connected facilities drive most of the growth at a projected 15 billion cubic feet per day, five times more demand growth than all other grid-connected sectors combined, with onsite generation adding up to 3.4 billion. The forecast accounts for announced projects that never complete, but if realized it would add roughly a million metric tons of daily greenhouse emissions and pressure gas prices that utilities may pass to ratepayers.

Filed from: TechCrunch , BloombergNEF .

Emerald AI, Google, and NVIDIA launch grid-flexibility alliance

Emerald AI, Google, and NVIDIA formed the AI Energy Management Alliance to advance data centers that dynamically adjust electricity use in response to grid conditions, through workload shifting, storage discharge, paired generation, or contingency response. The alliance’s technology-neutral principles cover ride-through, curtailment, and emergency obligations agreed before connection, standardized metrics and data sharing, and interconnection costs tied to actual system impacts. The announcement names no deployments, timelines, or capacity figures, so its practical effect is still to be shown.

Filed from: NVIDIA Blog .

Cornelis and Delos pitch open rivals to NVIDIA NVLink

At this week’s AI Infra Summit, Intel spin-off Cornelis Networks and newcomer Delos Data entered the scale-up networking race with open alternatives to NVIDIA’s NVLink, the fabric that lets hundreds of GPUs act as one accelerator. Cornelis introduced its Active Compute Fabric for open scale-up and scale-out networking alongside imminent 800 Gbps CN6000-series switches and NICs, while rivals such as AMD tunnel the emerging Ultra Accelerator Link protocol over standard Ethernet pending purpose-built switches. The pitches remain pre-deployment promises rather than shipping hardware displacing NVLink.

Filed from: The Register , Cornelis .

Agility unveils humanoid robot that yields to nearby humans

Agility Robotics debuted the Digit 5, its first humanoid designed to work near people without cages or separation barriers, using a safe-motion system that moves aside, stands still, or squats into a seated position as humans approach. The company says earlier Digits have logged more than 65,000 hours in warehouses and factories for clients including GXO, Schaeffler, Amazon, and Toyota’s Canadian manufacturing arm. Customers get early access in the first half of 2027 with general availability by year’s end, built at Agility’s retooling RoboFab plant in Salem, Oregon.

Filed from: Ars Technica , Agility Robotics .

Fluxnium claims a path to uranium from seawater

Stealth clean-tech startup Fluxnium claims polymer fibers licensed from US Department of Energy labs can extract uranium from seawater economically, hanging braided lines from ocean buoys for 30 to 60 days before recovering the uranium ashore without toxic tailings. The company closed a $7 million seed round led by Congruent Ventures with Constellation Energy, operator of the largest US nuclear fleet. Lab demonstrations ran above $200 per pound, more than twice Western supplier prices, and competitiveness rests on the founder’s unverified claim that larger fiber surface area has driven costs down.

Filed from: TechCrunch .

Meta lets AI agents set up WhatsApp Business messaging

Meta released a WhatsApp Business Tools MCP server that connects an AI coding agent such as Claude, Cursor, Codex, or ChatGPT directly to the WhatsApp Business Platform, replacing manual work across the Developer Console, Business Manager, API reference, and editor. The agent can create the business account, verify the phone number, register Cloud API access, build messaging templates, and monitor quiet failures such as payment or verification lapses. The move extends Meta’s existing MCP servers for ads and app configuration into business onboarding.

Filed from: TechCrunch , Meta developers .

AWS open-sources Pizza Bot, an email-style inbox for AI agents

AWS engineers open-sourced Pizza Bot, a local tool that manages AI agent work the way an email client manages messages, with finished tasks landing in an unread-style view and decisions awaiting the user shunted to action. The maintainers argue live chat assumes both parties are present and breaks the moment a task takes minutes, while a thread is a unit of work to return to. The project targets anyone juggling background agents that stall on permissions or produce reports needing review.

Filed from: The Register , AWS open source .

Infoblox ties casino-site sprawl to laundering and APT infrastructure

Infoblox reports tracking about 1.7 million Chinese-language casino sites facilitating illegal gambling, some of which double as command-and-control infrastructure for espionage and malware distribution while supporting North Korean money laundering and tax avoidance. The firm says major US hosting providers continue to host associated infrastructure, likely through stolen accounts in a practice known as infrastructure laundering. The findings come from the vendor’s own threat research and rest on its visibility rather than independent corroboration.

Filed from: The Register , Infoblox .

Nitter and XCancel shut down after X’s legal pressure

Nitter’s GitHub repository was archived into read-only status on September 11 and XCancel suspended service on Monday, ending the privacy-friendly front ends for reading X posts without an account. The closures follow an August cease-and-desist from X alleging unlawful API use and data scraping, after a brief September revival. Nitter stripped ads, tracking cookies, and JavaScript from public posts, which X had no interest in tolerating since it cannot monetize off-platform views.

Filed from: TechCrunch , Nitter .

Former Infosys chief’s AI startup adds $53 million to seed

Hang Ten Systems, founded in May by former Infosys CEO Vishal Sikka, added $53 million led by Temasek's Xora platform five weeks after a $32 million close, taking seed funding to $85 million with Aramco Ventures, Intel CEO Lip-Bu Tan, Micron CEO Sanjay Mehrotra, and Yahoo co-founder Jerry Yang participating. The startup advises large enterprises on AI strategy and builds production software with an in-house framework called Hobie for regulated industries. Deal and customer figures, including a multimillion-dollar contract signed within 25 days and work across 21 enterprises, come from the founders and are not independently verified, and valuation was not disclosed.

Filed from: TechCrunch .

From the Community

Apple Reference Image: A New Approach for Verified Photography

Apple introduced Apple Reference Image, an opt-in camera mode debuting on the main sensor of iPhone 18 Pro and iPhone 18 Pro Max that produces a securely timestamped image reflecting what the camera sensor captured. The sensor cryptographically signs pixel data immediately after capture, and Private Cloud Compute performs demosaicing, tone mapping, and compression in an environment Apple says it cannot inspect. Apple says the final signature is a composite post-quantum signature combining RSA-3072 and ML-DSA-87, with a revocation system that can invalidate individual photos or all photos from a specific sensor without exposing photographer identity.

Filed from: Apple Security Research .

Building a Linux GPU Driver for the M4 Mac Mini in One Month

Cody Ho and Niklas built a fully OpenGL ES 3.0 compliant GPU driver for the M4 Mac Mini and MacBook Neo in about a month, a process the author says normally takes years. The work involved reverse engineering the AGX firmware ABI and user-space components using hardware traces from a hypervisor, with the author describing the approach as a transparent, verifiably clean room effort that avoided looking at Apple binaries. The driver runs Minecraft at 200fps and runs Chrome and Firefox with working WebGL compositing, though the author says the code is not yet ready for end users.

Filed from: Cody Ho .

Mistral and Mozilla Bring Open Models to Firefox Smart Window

Mistral announced a partnership with Mozilla to power Firefox Smart Window, Mozilla’s AI browsing assistant, with Mistral models. Smart Window helps users make sense of complex searches and source information from their browser tabs, and Mistral says it will power the feature for users in France and North America, with the United Kingdom and Germany expected later this year. Mistral says conversations are not saved on Mozilla’s servers by default and that partners agree to zero data retention, though the announcement did not detail which specific models are used.

Filed from: Mistral .

Autistici/Inventati Shuts Down After Terrorist Designation

The Autistici/Inventati collective announced it is shutting down and discontinuing all services, saying it was designated a global terrorist organization and that continuing to operate would endanger users and their communities. The Italian collective has run privacy-focused email, mailing lists, and web hosting for roughly 25 years, and says every day online since August 26, 2026 has been a victory. It plans to publish instructions for backing up blogs, mailboxes, and websites, and warns that the autistici.org domain was made unreachable without notice and that further disruptions may follow.

Filed from: Keep the Internet Free .

TypeSafe AI opens early access to Jev, a model that returns decisions instead of text

TypeSafe AI, founded by former OpenAI researcher Diogo Almeida, has opened early access to Jev, the first of a model class it calls System One Models. Jev does not generate strings; it returns type-safe structured values with calibrated probabilities for classification, routing, scoring, and extraction inside ordinary software. TypeSafe says Jev is roughly two orders of magnitude faster and cheaper than comparable LLMs on these tasks, but its published evaluations were built by its own capabilities team, so the performance claims await independent testing.

Filed from: TypeSafe AI .

Strix finds a live admin GitHub token in a public Baseten container image

Security vendor Strix reports that its autonomous agent pulled a publicly accessible image from a Baseten Harbor registry and found an active GitHub personal access token for the account basetenbot recorded in the image’s build history. The token dated to a March 2023 build but still worked in July 2026, carrying admin and push rights on Baseten’s main product repository, its GitOps cluster repository, and its Homebrew tap, plus read and write access to other private repositories. Strix says Baseten confirmed the issue as critical, made the Harbor project private, and rotated the token the following day.

Filed from: Strix .

Google launches Gemini 3.8 Live voice models with parallel reasoning

Google DeepMind released Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking, two voice-first models handling near real-time visual context, background tool and API calls, and mid-conversation switching across 97 supported languages. The Extended Thinking variant reasons and speaks simultaneously, narrating progress while multi-step work runs in the background. Google says Extended Thinking took the top spot on Artificial Analysis’ Speech to Speech Quality Index, while Live placed second in the Speech Agent Arena, with all generated audio carrying a SynthID watermark; benchmark claims are Google’s own until independently reproduced.

Filed from: Google .