Security firm Phorion reports that versions 2.1.312 and 2.1.313 of the Claude Code for VS Code extension on the Cursor Marketplace contained an injected SSH backdoor. Cursor has removed both versions. →
Hackers drained around $6.3 billion from Binance wallets, most of it in about 20 minutes, in what appears to be the largest crypto theft ever. Three tracing firms link it potentially to North Korea; roughly half the funds are frozen. →
Northflank's status page shows an open security incident with major outages across seven components. The company says environment variables may have been decrypted, and users report their AI API keys were used without permission. →
An open-source project runs the NES's 6502 CPU, its RAM and a game cartridge inside an ordinary Qwen3.5 checkpoint whose weights were computed by a compiler, not trained. It is slow, at 150 to 290 times slower than a real console. →
A security researcher's third writeup on Azure API Connections (part of Azure Logic Apps) describes further ways to reach another tenant's backend, including Key Vault and SQL access, and says these findings alone netted $200,000. →
Apple says Apple Intelligence reaches China in an iOS 27.3 beta on November 4, with a full release in December. The launch covers earlier features and Alibaba's Qwen, not the new Siri. →
Nvidia is reportedly discussing a guarantee on billions in debt for Nscale's planned West Virginia data center, built for Anthropic under a $44.6 billion contract for the first building. The talks were reported by The Information and corroborated by Bloomberg. →
Four of the biggest AI companies say they will honor a shared "policy token" in robots.txt, so publishers can set AI training preferences once. Meta voiced support but did not say whether it will honor it. →
Google says it booted the TPUs on its Project Suncatcher satellite and that the on-board Gemma 4 model answered a prompt to write a haiku about space. Radiation and thermal tests come next. →
Tangled, a git forge built on the AT Protocol, has released a redesigned Svelte UI, a Rust indexer called bobbin, a Rust knot server, Organization accounts and a reworked stacked-PR flow. More launches follow over the next few weeks. →
A new paper turns the Platonic Representation Hypothesis into a theorem for JEPAs and reports that a 109M-parameter text JEPA aligns with image encoders better than language models up to 72B parameters. →
A new paper pairs a compact 3D Gaussian memory with a video diffusion transformer to keep generated worlds spatially consistent in real time, at 9 FPS. The authors report up to 38.6 times fewer 3D primitives than the most efficient prior 3D memory. →
OpenAI has reorganised the Model Spec, its public behavior guidelines for AI models, around five values and added firmer correctability rules for agents. The Decoder notes that following the rules is another matter. →
In a guest post on virtuallyfun.com, Rairii says a port of classic Mac OS to Nintendo's ArtX-chipset PowerPC machines has reached the desktop for the first time. The post walks through the boot patches and drivers it took; the project is not ready for public release. →
Alphabet missed estimates on its third-quarter results, and CNBC says the stock fell about 13% on Thursday, the steepest drop on record. Search revenue grew 12% to roughly $63 billion while Google Cloud revenue more than doubled. →
Youtu-Parsing-Omni is a single 5B-parameter model that turns documents, images, audio and video into structured records. It reports 96.96 Overall on OmniDocBench v1.6, and code and weights are public. →
Reddit CEO Steve Huffman says the company lost more than one million Google search referrals per week in October and blames AI Overviews and AI Mode. Shares dropped over 25 percent despite strong earnings. →
After Alphabet, Amazon, Meta and Microsoft outlined capex plans of more than $740 billion this year, TechCrunch says investors are worrying about negative free cash flow and debt. The Nasdaq fell 2.9% on Friday, and Alphabet had its worst day on record on Thursday. →
Blackstone, anchor investor in a $63 billion Anthropic chip-financing deal, reportedly told investors it expects a double B rating after the IPO. It had earlier predicted investment grade before 2027. →
On the Cheeky Pint podcast, OpenAI CEO Sam Altman called it "terrible" for the world's limited DRAM to sit in personal computers, used about five percent of the time, as the RAM shortage drags on. →
A new paper says the standard multi-token prediction loss carries a noisy gradient residual that gets worse with scale. Soft-MTP drops it with a one-line change and reports gains from 167M to 1.2B parameters. →
A new paper adds a moment-matching loss to distribution matching distillation for one- and few-step text-to-image models, using the generator's own EMA copy instead of an external encoder. At two network evaluations it reports 94.25% (Chinese) and 90.20% (English) on LongText-Bench. →
Microsoft Research describes a platform that generates fully synthetic company environments, grounded in public data, to test Copilot without touching customer data. It reports a 39% gain in a composite quality score and estimated savings of about $100,000 per dataset. →
In a Twitter/X thread, a FurtherAI author describes automating each step of shipping software over nine months, with self-reported gains, and argues the new bottleneck is deciding what to build. →
Firezone has open-sourced sanstun, a userspace WireGuard state machine in Rust that separates packet bookkeeping from encryption so crypto can run on other threads. It replaces the company's boringtun fork, which is now deprecated. →
Cisco president Jeetu Patel says he has handed a 27-year-old, Padmanabhan Krishnamurthy, unlimited AI tokens to refactor the company's legacy code, with a goal of 98% refactored within 24 months. →
Just over a week after Granite Switch 4.1, IBM released a backwards-compatible 4.1.1 patch with a new multilingual speech-to-text model and an improved Decision adapter, trained without touching the base model. →
Wired says Anthropic's IPO filing admits the company lacks hands-on lab and drug-discovery experience and softens its language on a Claude-found enzyme, after months of bolder talk from CEO Dario Amodei. →
Import AI 478 covers a startup's roughly $135,000, week-long agent run that produced about 600k lines of Lean, Claude Opus 5.5 reaching 18.5% on ProgramBench, and a one-researcher pretrained model from Mirendil. →
OpenAI says it is strengthening violence prevention in ChatGPT with expanded safeguards, earlier interventions and a new commitment to public safety reporting. →
A preprint called Glance gives a RAG generator a small decision adapter and pointer head so the same 9B model scores candidate passages, replacing a separate reranker. It reports small gains over bge-reranker-v2-m3 on two retrieval sets; the authors call the results initial. →
A new paper, A-RSI, decides when a self-improving multimodal model should get fresh training tasks by watching its own training logs. On an eleven-benchmark suite, the performance average rises from 57.77 to 61.05 with 800 fresh tasks. →
A new training-free decoding method, LOWN, checks at each step whether a diffusion vision-language model needs the image at all. On LLaDA-V it keeps 99.6% of the score at 35.3% of the decoding FLOPs on average. →
A new simulated benchmark built on an Isaac Lab digital twin of the SO-101 arm tests language-conditioned spatial tasks. The best learned policies score near 61%, while an LLM agent that writes code reaches 83.2%. →
A new benchmark, Investor's Last Exam, scores AI financial agents on equity research using a frozen information set. GPT-6 Astra and Claude Opus 5.5 lead at 68.1 and 66.9 out of 100, but Astra manages only 52.5 on Forecast & Model. →
A new paper presents EditAgent, an image-editing agent that routes steps to general and specialized editors and gives its judge a verification record for each candidate. The authors report gains for two open-weight editors on three benchmarks. →
A new benchmark built from real user coding sessions finds that the strongest of ten model endpoints passes 69.9% of criteria per task on average, but fully satisfies only 26.4% of tasks. →
A new paper introduces SpecSlot, a pixel-space rectified-flow Transformer in which only the lowest-frequency slot of each anchor joins global attention. On ImageNet-512 it improves FID by 2.84 over a FLOPs-matched JiT control. →
Intel shares fell around 28% on Friday after the company guided to fourth-quarter revenue below analyst expectations. CNBC reports the drop cut the value of the US government's stake by $12 billion. →
Dartmouth provost Santiago Schnell resigned after an advisory committee found he did not disclose significant AI use in an academic paper and a Washington Post op-ed on AI cheating. →
A preprint trains small UI navigation agents with online reinforcement learning on live iOS apps. The authors report that their 8B agents beat SFT baselines and tie Gemini-3.1-Pro on real iOS applications. →
A new Sierra Club analysis says US power-sector CO2 emissions could be 42 percent higher in 2030 than in 2025 as fossil fuel plants are built for data center demand. The figure depends on how many planned projects get built. →
