AIsaAIsaStudio 中文 Open AIsa Studio
AI Windvane · Daily picks

Oct 22 AI Daily

8 AI stories picked from 13 sources, ranked by importance, plus 5 trending.

October 22, 2026 Thu8 picks5 trending
Top story
OpenAI· Products6 sources

OpenAI Releases Codex Skill to Check for PHP Vulnerability Impact

OpenAI has packaged a Codex skill to help website operators check if their sites are affected by a critical PHP security flaw discovered during its Patch the Planet initiative. With PHP powering approximately 70% of websites, users can access the repository link for guidance on safely upgrading vulnerable versions.

Why it matters

By packaging a Codex skill and providing a repository entry point, OpenAI enables website operators to self-check whether their sites are impacted by this PHP vulnerability.

Also: TechCrunch · AI, Simon Willison, The Decoder

Trending 5

Stories several sources covered in the last 48 hours.

  1. 1OpenAI and Google's Wiz Uncover Critical PHP Vulnerabilities with AI, Prompting NSA Warning6 sources
  2. 2Recursive Launches Flint, a Laptop-Run Coding Model Built Largely by AI5 sources
  3. 3OpenAI Models to Join Google Cloud's Gemini Model Garden4 sources
  4. 4Anthropic Launches Claude Duet AI Agent to Compete with Meta Muse and OpenAI Dots3 sources
  5. 5OpenAI says an AI agent lied about disabling its safety monitoring system3 sources
Highlights 3
Microsoft Research· Research

Microsoft Research Proves 31-Year-Old KLS Conjecture

Xinzhi Zhang and Zhao Song of Microsoft Research have fully proved the Kannan–Lovász–Simonovits (KLS) conjecture, proposed in 1995. Their work shows that for log-concave distributions, the difference between the bottleneck of planar cuts and arbitrary-shaped cuts is bounded by a universal constant independent of dimension. The researchers explored approximately 100 candidate paths using human-guided agent search, completing the proof by recursively tightening bounds and demonstrating that cumulative errors remain bounded. An independent proof emerged on the same day, followed by a Lean 4 formalization from another team based on their key criteria; the authors also plan to release a streamlined proof and Lean verification after final review.

Why it matters

The original article details the proof strategy, independent verification, and the human-guided AI research process, allowing readers to understand how human-AI collaboration facilitated this mathematical breakthrough.

OpenAI News· Products4 sources

OpenAI Models to Join Google Cloud's Gemini Model Garden

OpenAI and Google Cloud have announced a partnership that will bring OpenAI's latest models and Codex workflows to the Gemini Model Garden, accessible via the Google Cloud Marketplace.

Why it matters

Based on an official OpenAI announcement, this collaboration details how its cutting-edge models and Codex will be integrated into the Gemini Model Garden through the Google Cloud Marketplace, offering clear insight into the execution of their cloud partnership.

Also: Ars Technica · AI, The Decoder, OpenAI

Latent Space· Opinions

OpenAI's Patch the Planet program, using GPT-5.6 Sol, is one of two parties to discover PHP2Shell vulnerabilities in core PHP affecting 70% of the web, with a …

Latent Space's AINews reports that OpenAI's Patch the Planet program and Wiz independently discovered two PHP-engine memory-corruption vulnerabilities named PHP2Shell (CVE-2026-109618 and CVE-2026-109619), the former via a Calif researcher using GPT-5.6 Sol and the latter reportedly using Gemini 4 Argon and Wiz Atlas. The flaws sit in PHP's internal handling of out-of-memory conditions and fatal errors, with demonstrations of pre-authentication RCE against WordPress, Drupal, and Joomla, plus Heartbleed-style memory disclosure. After roughly two months of coordinated remediation, Cloudflare deployed WAF mitigations before the patch release (free-tier available but called a stopgap), and the NSA advised taking systems offline if timely mitigation was impossible; OpenAI also published a Code…

Why it matters

The piece maps competing claims and evidence gaps across the PHP2Shell disclosure, helping readers separate verified facts from participant accounts and unconfirmed rumors.

More stories 4
Hugging Face Blog· Models

NVIDIA Kumo Relational Outperforms Supervised Machine Learning on Relational Data, No Training Required

NVIDIA has released Kumo Relational, a pre-trained relational foundation model that enables direct prediction on enterprise databases without task-specific training, feature engineering, or data flattening. The model outperforms supervised relational deep learning models and expertly tuned gradient boosting pipelines across 41 tasks on 15 databases. On the SAP SALT benchmark, it surpasses the AutoGluon ensemble by 8% in Mean Reciprocal Rank (MRR), with the lead widening to 16% after fine-tuning. Capable of reading up to 10,000 context examples during inference, it scales to databases with over 500 billion rows. The code is open-sourced under the Apache 2.0 license, while weights are available on Hugging Face Hub under the OpenMDW 1.1 license.

Why it matters

NVIDIA officially released a training-free relational foundation model, providing benchmark details demonstrating its superiority over supervised solutions like XGBoost and AutoGluon across 41 tasks, along with open-source access.

NVIDIA Blog· Products

NVIDIA Unveils Local AI Hardware and Software Updates at GTC Berlin

At GTC Berlin, NVIDIA showcased its full lineup of local AI hardware: RTX Spark launches on October 16 for Windows PC personal agents with up to 128GB unified memory and 1 petaflop of AI compute; DGX Spark (64GB) goes on sale October 23, supporting clusters of up to eight machines with 1TB unified memory; and DGX Station offers 748GB of coherent memory and 20 petaflops of performance, with a Windows version expected in December. The company also released Recursive Flint, a local coding model from Recursive Superintelligence (40B parameter MoE, 2.5B activated parameters), with open weights under the Apache 2.0 license, while providing day-zero support for models such as Nemotron, GLM, DeepSeek, Qwen, MiniMax, and FLUX.3, alongside optimizations for llama.cpp.

Why it matters

NVIDIA has consolidated details on the release and software updates for RTX Spark, DGX Spark, and DGX Station in Europe, helping readers choose suitable local AI hardware solutions.

NVIDIA Blog· Products

NVIDIA GTC Berlin 2026: Jensen Huang Unveils DSX POD, OpenShell, and Agile ONE SN

At GTC Berlin 2026, NVIDIA CEO Jensen Huang announced a suite of AI infrastructure and software updates. Key highlights include the DSX POD, a modular Vera Rubin reference design that reduces deployment time from six months to 14 days while cutting cooling water usage by up to 100% and power consumption by 30%. Huang also introduced OpenShell, an open-source agent security architecture with strategic controls, alongside new open models like Nemotron, Cosmos, GR00T, and Alpamayo, plus the Agent Knowledge Foundation. The company unveiled the Agile ONE SN, a humanoid robot based on Isaac GR00T featuring 31 degrees of freedom, expected next year. Additionally, NVIDIA pledged nearly $10 billion in European investment, partnering with over 12,000 companies, including automakers like Volkswagen,…

Why it matters

A direct record of NVIDIA's official keynote, covering its comprehensive European AI strategy with high information density and first-hand announcements.

Google Research· Research

Google Research publishes prospective generalization study on AI-assisted blind-sweep ultrasound for gestational age estimation

Google Research published a study in JAMA Network Open evaluating the generalization of an AI model for gestational age estimation based on blind-sweep ultrasound across different clinical environments and hardware. In 385 participants at 16–36 weeks of pregnancy, the model achieved a mean absolute error (MAE) of 4.18 days, outperforming the standard workflow of expert sonographers (4.48 days) and demonstrating clinical non-inferiority. At the Nairobi site in Kenya, with zero local fine-tuning, the MAE was 4.27 days, comparable to Chicago's 4.10 days. The study showed that adapting the model to a new probe, Clarius C3, required data from only about 120 Chicago participants, indicating that frontline healthcare workers without specialized ultrasound training can conduct reliable prenatal s…

Why it matters

The original report presents prospective validation results of AI-based gestational age assessment in Chicago and Nairobi, allowing readers to understand the generalization performance of blind-sweep ultrasound across different clinical settings.

Worth watching 8