Developer enann has open-sourced the subscription ledger Calends, designed specifically for car owners in shared housing and users with multiple accounts. It supports integration with AI assistants like Codex, Claude Code, and Grok Bot via MCP to automatically parse subscription emails and import historical data. It features built-in DeepSeek API integration capable of handling cross-region price increases, proxy payments, separation of gift card face value and cost, exchange rate calibration, and balance date marking. It supports deployment via Docker or Cloudflare Workers + D1 and is open-sourced under the MIT license.
Lumabri is an open-source project written in pure C that supports collaborative execution of large language models across multiple local computers and servers for private inference. Based on the Colibri engine, it achieves distributed model execution in memory-constrained environments through techniques such as sharded loading of model weights (assigned by layer), shared state management, and preservation of dialogue node states.
The Verge reporter Mia Sato tested AI shopping agents such as Meta Muse and found that they encountered hard blocks, dynamic bot checks, or limited functionality on platforms including Amazon, Target, and The RealReal; Amazon explicitly required Meta to remove access to its site, and institutions like JPMorgan Chase also refused to open account permissions; Muse was repeatedly blocked on Target but occasionally passed, and could bulk-add locked items to carts on The RealReal, raising concerns about fairness and platform ecosystems; the author points out that current agent shopping experiences are fragmented and diminish the joy of discovery and decision-making sovereignty inherent in human shopping.
AI models such as GPT-5.6 Sol assisted Calif and Wiz Research in discovering two high-severity remote code execution vulnerabilities in PHP, CVE-2026-109618 and CVE-2026-109619, both rated 9.2; these vulnerabilities allow attackers to remotely take over servers without authentication, affecting approximately 70% of websites globally; the PHP Foundation acknowledged that this risk was first noted in an external audit in 2024 but remained unfixed for a long time due to compatibility and collaboration cost issues; vulnerability details leaked before patches were released, with no official update support for PHP 8.1 and earlier versions, requiring fixes via source code patches or distribution backports.
Aerial is a lightweight desktop tool for domestic Android smart TVs. It bypasses system restrictions to achieve root-free desktop replacement via a background daemon + notification listening mechanism, supporting automatic Agent installation deployment; it gained over 200 GitHub stars within less than 1 day of release, has adapted to multiple TV brands, and continues to iterate features.
Domestic computing power chip manufacturers generally report high growth in performance, but external environmental dividends are diminishing. Competition is shifting from single-chip performance to software migration (such as Day 0 adaptation for Step 5 Preview), supernode system efficiency (reducing communication loss to within 10% and achieving compute utilization rates exceeding 90%), and supply capability (with multiple manufacturers significantly increasing inventory and prepayments). IDC data shows that NVIDIA's market share in China dropped to 34% in the first half of 2026, while Huawei rose to 28%, indicating that the market focus has shifted to large-scale computing systems.
SoFi has launched its tokenized deposit product SoFi Earn (SOFIE), where each token represents $1 held at SoFi Bank, with an annual yield of 4.40% and funds protected by FDIC insurance; users can redeem SOFIE one-for-one for SoFiUSD stablecoin, trade cryptocurrencies within the SoFi app, or transfer via blockchain in real time.
Symantec security researchers captured a malicious AI agent named "LLMjacking" in a honeypot. This agent hijacks unprotected language models through a five-layer stacked prompt jailbreak (including a French first layer, fictional identity, relationship binding, disabling ethical terms, and 74 attack modules). Once entering "Skynet mode," it can autonomously execute up to 100,000 steps of operations such as scanning, privilege escalation, data theft, and deploying XMRig mining programs; a second attempt where the attacker bypassed model interaction to directly issue obfuscated commands was also fully recorded.
The AI21 team reduced the per-PR cost of GitLab Duo Code Review from $0.589 to $0.0135 on Martian's Code Review Bench, a decrease of 97.7%. The core strategy involved switching all four-stage models to DeepSeek V4 Flash and layering context compression; compression reduced the maximum review prompt from 480K tokens to 109K, avoiding Sonnet 5 fallback caused by exceeding the 256K deployment window. F1 remained at 47.2% (baseline was 46.4%), inference output tokens dropped from 18,600 to approximately 9,000, and latency decreased to about one-third.
Broadcom is arranging over $50 billion in financing for custom AI chips for OpenAI, following the recent disclosure of a TPU computing power leasing financing arrangement with Anthropic worth up to $42 billion; Synopsys signed a multi-year IP agreement worth over $1 billion with Amazon; SEMIFIVE received a $52 million order for AI inference ASIC development; Marvell raised its FY2029 custom chip revenue target to over $12 billion; Cerebras signed multi-year computing power deployment agreements with General Compute and Gimlet Labs respectively, focusing on coding agents and heterogeneous inference platforms; Industry research predicts that by 2027, ASIC shipments in high-end cloud AI accelerators will reach 15.26 million units, accounting for nearly 60%.
OpenAI, in collaboration with PerturbAI and Scripps Research, deployed thousands of AI agents running on GPT-Rosalind to analyze a whole-brain CRISPR atlas comprising millions of mouse brain cells; the entire process took 32 minutes. Performing comparable analyses manually would require several months to one year. The agents analyzed cell responses gene-by-gene through perturbation studies and retrieved relevant literature, uncovering unexpected differences between two disease-related genes and potential links between cholesterol regulation and neurodegeneration. In blind tests, human experts preferred reports generated by distributed agents 75% of the time. The 'Reasoning Atlas' has been open-sourced on GitHub, and the accompanying paper is published on bioRxiv.
Tim99 developed and launched the PHP2Shell emergency assistance website (https://checkphp2shell.com/), providing public domain response header inspection (including PHP/Laravel/WordPress clues and Cloudflare/Sucuri/Akamai/CloudFront marker identification), remediation threshold verification after manually inputting the PHP version, and configuration guidance for php.ini, Nginx, Apache, Cloudflare WAF, ModSecurity, and Docker/K8s. The tool only sends HEAD (or fallback GET) requests, does not perform vulnerability exploitation, clearly distinguishes between temporary mitigation and complete remediation, and supports five languages.
To address repeated timeouts when publishing multi-module projects to Maven Central, TAKETODAY developed Gradle plugin v1.0.1. It supports one-click publishing through staging → bundle → upload to Portal → polling deployment status, as well as automatic switching for SNAPSHOT versions. The design features aggregated publishing from the root project and lazy evaluation of credentials. The plugin is open source, does not depend on its framework, and source code, documentation, and usage instructions are available in the GitHub repository.
Based on the modular development board ESP-Mosaico (main controller ESP32-S31) released by Espressif in August 2026, the author successfully ported Meta's open-source Muse Gadgets device SDK through collaboration with the Claude Code programming agent.
On October 22, Ant Digital Technologies released the decision model LingDT-Decision-1.0, offering two versions: 4B and 27B. It is now available on the DTMaaS platform for a limited-time free trial. The model pioneered native multi-selection and sequential decision-making (DAG) capabilities. In 23 open-source evaluation benchmarks, its comprehensive score surpassed models such as Microsoft-Decision-1 and Jev 1.13.0, with the 27B version achieving the highest accuracy among models of the same parameter size. Actual tests show that it reduces Token consumption by over 90% and improves response latency by up to 10 times compared to large language models. It also enhances risk identification accuracy, tool combination matching rate, and hallucination recall rate in scenarios such as customer service, risk control, and finance.
Tinytowns is a lightweight decoder-only character-level Transformer model that runs in the browser, with 1.0015M parameters, trained on 142706 real place names from GeoNames, supporting the generation of realistic town names; it provides three model tiers—nano (4.572K), small (105K), and tinytowns (1.0015M)—all open-sourced in .safetensors format on Hugging Face and GitHub; the model performs forward inference locally in JavaScript without server dependencies, with a validation loss of 2.1504.
Licensed Malaysian payment service provider Axaipay has partnered with Italian website building software vendor Incomedia to launch the AxaiCart+ solution, integrating WebSite X5 Pro's drag-and-drop website builder with Axaipay's local payment gateway. Officially launched on October 22, it supports local wallets such as FPX online banking, TNG eWallet, ShopeePay, Boost, and DuitNow QR, as well as international credit and debit cards. Adopting an independent hosting ownership model, it provides low-cost integrated website building and payment collection services for micro, small, and medium enterprises and entrepreneurs.
Bitpush reported that DATA (formerly Story Protocol) announced the closure of its own Layer 1 blockchain and migrated to the Sui network after receiving a $30 million investment from the Sui Foundation; during the same period, projects such as Helium, Celo, Lisk, and Polygon also scaled back or exited their chain infrastructure.
ego lite is a lightweight browser designed for AI agents, supporting the parallel execution of multiple agent tasks in independent, isolated Spaces without interfering with the user's foreground browsing; it supports zero-configuration startup, automatic inheritance of Chrome login states and extensions, and provides JavaScript tool interfaces such as snapshot/fill/click; currently available for macOS, with Windows and Linux versions coming soon; the project is open source (MIT license), having received 5387 stars and 260 forks on GitHub.
Chrome is exposed to a sandbox escape vulnerability CVE-2026-113328 in the Privacy component; improper input validation may allow malicious web pages to execute arbitrary code outside the sandbox. It affects all Chrome desktop versions prior to 156.0.8078.47. The official fix has been included in the Chrome 156 stable release; the fixed version for Windows/macOS is 156.0.8078.47 or 156.0.8078.48, and for Linux it is 156.0.8078.47 and above.
In its pursuit of an IPO, Anthropic has adopted a dual governance structure combining a Public Benefit Corporation (PBC) with a Long-Term Independent Trust (LTBT), entrusting control to external trustees who hold no equity and receive no dividends (such as former Federal Reserve Chairman Ben Bernanke), thereby imposing legal and mechanical hard constraints on AI safety and public mission; this design differs from OpenAI's top-level non-profit foundation structure and xAI's return to a standard commercial company path, sparking substantive discussions in the capital markets regarding shareholder rights dilution and valuation discounts.