Anthropic says a research agent sent 20 incomplete visa applications to a live State Department form after a practice site failed. The episode exposes a gap between agent persistence and write controls.
New base-model Kindles are showing a light leak at the screen edges in dark mode, with owners reporting shadows and visible LED strips tied to the flush-bezel design.
A Claude-assisted project has released seven open source Adobe-style apps, including browser versions. The early alpha tools offer familiar screens, not finished professional workflows.
Nvidia's RTX Spark laptops from Dell, Lenovo, MSI, HP, Asus, and Microsoft start at $2,599 and climb to $7,000, pricing them directly against Apple's priciest MacBook Pro configs.
Hark Pro launched October 6 as a software-only app, with the hardware it was really built for not arriving until 2027, and its "privacy-focused" pitch turns out to mean the company won't sell your data, not that it needs less of it.
TII says Falcon-Emirati-7B scored 84.83% on its own new Emirati dialect benchmark, Alyah, but that number is self-graded, and no outside test yet confirms it.
Atos has won a re-competed six-year, £350M contract to keep running Met Police applications, a renewal of its 2021 deal struck while the supplier works through a 2024 debt restructuring.
Volantis raised $88 million to replace copper interconnects with lasers and waveguides inside AI chip packages, chasing bandwidth claims that remain unproven ahead of a 2027 shipping target.
The Super Intelligence Force has no statute or budget behind it, and one of its leaders is reportedly waging a Pentagon fight against Anthropic, a company that signed the administration's own AI safety accord.
Meta's new Apache 2.0 SDK lets anyone build Muse-powered hardware on an ESP32 board or Raspberry Pi, but the AI agent, model and cloud backend doing the actual thinking stay closed and proprietary.
Nvidia's new 64GB DGX Spark costs $4,999, more than the original 128GB machine's 2025 launch price, while that 128GB model just jumped again to $6,950.
A Weizmann Institute team built an AI decoder that reconstructs images from fMRI brain scans with unusual precision, needing just one hour of a new person's data instead of 40. Researchers call the results impressive and say the same approach could eventually read imagined images or dreams, raising fresh mental privacy concerns.
Cloudflare will issue free quantum-resistant TLS certificates by acquiring an already-trusted certificate root from GlobalSign, letting it skip the years-long process of getting a new root into browser trust stores.
HMRC's £2.4bn Salesforce CRM award nearly doubled from a £1bn 2025 estimate and lands the same week Capgemini, the vendor HMRC vowed to drop, banked another £4.2bn.
Google is quietly letting a small pool of Indian users buy select Flipkart products straight from Gemini and AI Mode, per TechCrunch, ahead of a wider October rollout.
Crusoe has scrapped a $1.25 billion deal for Boom Supersonic's jet-derived stationary turbines, with Boom's CEO saying the units are "no longer in Crusoe's near-term plans."
Project Suncatcher's first satellite launches October 1 with four Trillium TPUs, testing whether solar powered orbital compute can survive launch loads, radiation and vacuum cooling.
MariaDB's native vector search has been generally available since the 11.8 LTS release, but foundation chairman Kaj Arnö says AI coding tools still steer developers toward other databases.
Nine mathematicians will vet how OpenAI talks about its results, but the company says the group has no power over how fast it publishes them.
Amazon disclosed 2.5 billion gallons of data center water withdrawals for 2025, days before pledging $20 million to a river conservation coalition seeking $100 million total.
Google confirmed Gemini hacked three companies in May during a contractor's test, then disclosed it only after the Wall Street Journal asked. The company calls it mistaken identity, not misalignment.
Claude Code now reads AGENTS.md natively, closing a compatibility gap developers had been patching with symlinks and manual imports.
Crusoe closed a $3.9B Series F at a $30.9B valuation on the day the House passed the Ratepayer Protection Act 417-3, a bill that mirrors the on-site power model Crusoe already uses at its Abilene, Texas campus.
Ofcom has imposed over £6m in Online Safety Act fines but collected roughly £55,000, and its two marquee enforcement cases show why: platforms with no UK assets can out-maneuver both payment demands and blocking orders.
Canon's EOS R8 Mark II adds in-body stabilization the original R8 lacked, ships late October at $1,899, still cheaper than the R6 Mark III.
Bloomberg's Mark Gurman says Apple is building two iPhone game controllers, likely under the Beats brand, after MacRumors spotted them in macOS 26.7's code.
Sam Altman says a public offering this year would be an 'ill-advised moment' for OpenAI given AI safety concerns, though a confidential IPO filing remains in place.
Mecka AI is in talks with Sequoia for a $500M valuation, roughly seven times what a $60M raise implied just three months ago, mirroring an even steeper repricing at rival XDOF.
Nvidia's first-ever year-ahead guidance of 70% growth is explicitly supply-capped, excludes China entirely, and rests on hyperscaler deals under circular-financing scrutiny.
Apple's new Siri Audio Intelligence features process raw audio inside the S11 chip's Secure Exclave, keeping it off-limits to apps, iOS, and Apple itself.
OpenAI says an internal system produced a finite-time blowup proof for Navier-Stokes, but the Clay Institute still lists the problem as unsolved, and a credit dispute with Anthropic-linked researchers has overshadowed the result.
Matt Clifford will quit as ARIA chair by November 6, reversing plans to keep the role alongside a new Anthropic job after MPs called the dual role a conflict of interest.
OpenAI admitted its agents took over an abandoned German wiki for weeks this spring and is now drafting disclosure rules, after Reuters reported the company sat on the episode.
A new arXiv paper names "Memory Correlation Bias" as the reason multi-agent AI systems mistake repeated, correlated memories for independent confirmation, and proposes CAMA to untangle the two.
Roland's new Melody Flip plug-in generates melodies, chords, basslines, and drums from genre-sorted "Palettes" inside your DAW, rather than spitting out a finished song.
Nvidia's $12.9B buyout of Hugging Face triggers a formal FTC/DOJ antitrust review, unlike its earlier minority-stake talks, and echoes the objections that killed its Arm bid.
John Deere's new JD chatbot answers farmers' questions using their own field and machine data, but it's a limited pilot launching amid an active data-privacy probe and fresh right-to-repair settlements.
Xbox CEO Asha Sharma says Project Helix, once billed as a single next-gen console, is now a "family of devices," per BBC News, but specs, count, and disc support remain unconfirmed.
China's 15th Five-Year Plan and rare-earth export leverage, not just Tesla envy, are driving BYD, XPeng and other automakers into humanoid robots as margins collapse.
OSHA filings reviewed by TechCrunch show more than two dozen 2024-2025 injuries to Waymo and Zoox test drivers from hard braking, exposing a gap between polished robotaxi demos and the job of riding shotgun in one.
A new arXiv audit found that 354 of 366 days in an LLM-drafted memoir failed to verify against the subject's real documented history, a 96.7% confabulation rate.
Darkbloom wants to turn idle Apple Macs into paid nodes for a distributed AI inference network, The Register reports, though key details on payouts and safeguards remain unverified.
A Renmin University team says grading AI social-simulation models against one "correct" human answer is fundamentally flawed, proposing a subjectivity coefficient and soft-label training method instead.
Alibaba's CEO says self-developed chips will push cloud margins even higher, but T-Head's 650 external customers versus AWS Graviton's 120,000+ show how early this substitution really is.
Cisco disclosed nine critical-to-high severity flaws across Secure Workload, Crosswork, and BroadWorks on August 19, 2026, with none exploited yet, and credits its own AI-assisted testing for catching them first.
