For people who build with AI.

Artificial intelligence is automating part of reverse engineering by analyzing programs, identifying features and generating code without access to the original source code. MIT Technology Review Brasil cites developer Brandon Thomas, who used Anthropic’s Claude to create free, open-source versions of Adobe applications available on the ArtCraft platform. Tests by Creative Bloq found important gaps compared with the original products, including lower precision and missing features.

In another project, Maurice Heumann says he reconstructed 99% of the functions of Call of Duty: Modern Warfare 2 from executable files, with 83% matching the original byte for byte, after about three months. He says the work used up to 17 agents simultaneously and consumed more than 500 billion tokens. The case also revealed code that appeared correct but contained important errors, underscoring that AI can speed development but still requires testing, supervision and technical expertise.

Sources: MIT Technology Review Brasil

Read the original story ↗

In a post, Hugging Face’s PyTorch team says PyTorch 2.15 adds FP8 to MPS, using half the space of BF16, along with new paths to speed up language models, mathematical operations, audio, video and training. The team reports that mixture-of-experts (MoE) models became 5–6.5× faster to decode; on an M3 Ultra, Qwen3.6-35B-A3B rose from 9.4 to 61.3 tokens/s, while softmax over the vocabulary became about 8× faster.

The team also says Wan2.2’s video VAE decoding now peaks at about 8 GiB of memory, down from 270 GiB, and that a fix improved Qwen-Image-2.1 image quality on MPS. PyTorch 2.15 requires macOS 15 or later and Python 3.11 or later.

Sources: Hugging Face

Read the original story ↗

Hugging Face has released AnyLanguageModel 1.0, a Foundation Models-compatible API that lets developers switch models and providers without changing the core session code. The tool now offers tool calling, guided generation, and image support with all compatible providers, as well as support for system models, open-weight models through llama.cpp, MLX, and Core ML, and services from OpenAI, Anthropic, Google, Ollama, and Hugging Face Inference Providers.

The update also adds Linux support and comes with swift-huggingface 1.0 and swift-xet 1.0, which are aimed at accelerated model downloads in Swift apps. Hugging Face highlights smaller models such as Gemma 4 E2B and LFM2.5 2.6B for use on an iPhone, while Qwen3.8 27B requires a Mac with 32 GB of memory or more.

Sources: Hugging Face

Read the original story ↗

Hugging Face presents TorchTPU, a new PyTorch backend for TPUs, Google’s AI accelerators. It lets developers use regular PyTorch code and send models and inputs to the “tpu” device, with support for Transformers, Diffusers and Kernels; the package also includes Pallas kernels, distributed training and torch.compile integration.

According to Hugging Face, TorchTPU removes the need for Optimum TPU to run Transformers and Diffusers models on TPUs. Transformers can load models directly onto the chips, while Diffusers offers eager and compiled execution modes, as well as sharding large models across multiple chips through TensorParallelConfig. Installation on a TPU VM requires a single pip command.

Sources: Hugging Face

Read the original story ↗

Amazon Web Services released TorchNeuron, a tool that lets AI startups adapt applications written in PyTorch to Trainium chips with a configuration change and no code rewrite, Amazon says. The company says about 85% of AI applications use PyTorch, whose integration with CUDA software helped make Nvidia chips a default choice. TorchNeuron is adapted from PyTorch and can also be used by developers familiar with the project.

Beyond making the move to Trainium easier, Amazon says it will contribute changes to the main PyTorch project to remove code tied to specific chips. AWS says its cloud integration and using the same chip for training and inference can reduce costs; according to Colin Brace of AWS Annapurna Labs, Deepgram processes 70% more audio per dollar on Trainium while maintaining production-grade accuracy. That assessment comes from AWS.

Sources: TechCrunch

Read the original story ↗

OpenCode said attackers accessed an older system powering OpenCode Console, affecting 1.4% of the platform’s users. Some API keys stored by users were exposed, and developers reported unauthorized charges on OpenAI, Anthropic and Google accounts, including cases in the hundreds of dollars and some above $1,000. Anthropic and Google blocked keys they considered compromised; OpenAI said it worked with OpenCode to restore access and refund credits used in the attack. OpenCode said refunds are underway.

According to co-founder Dax Raad, the open-source app and command-line tool were not affected. During the response, AI models refused to analyze malicious code after triggering safety filters. Raad said later access to OpenAI’s Daybreak Blue program improved the investigation. The episode renewed criticism that safeguards can hinder defensive security teams.

Sources: TechCrunch

Read the original story ↗

AI security firm Norforge says it identified a hacker who used the open-source Cairn platform to coordinate largely autonomous agents. Norforge’s analysis says the agents tried to breach more than 1,200 websites and compromised more than 100; Reuters identified pages linked to John Deere, SharkNinja and Cummins among the targets. The attacks reportedly took about 23 hours each, and Norforge found more than 20 exposed Cairn instances that were likely malicious.

Xataka reports that the attacker’s instructions were fewer than 100 Chinese characters long and estimates, based on Snorre Fagerland’s analysis, that the campaign consumed about $5,000 in tokens — roughly $50 per compromise. Cummins told The Verge it is investigating a potential attack on a website operated by a third-party vendor; John Deere told Reuters there was no indication that sensitive information was accessed, while SharkNinja said the incident affected a standalone site and had been contained. Norforge is working to take down the malicious instances.

Sources: The Verge, Xataka

Read the original story ↗

Hugging Face announced TorchNeuron, a PyTorch plugin that adds a native backend for AWS Trainium. The company says the tool supports eager execution, torch.compile and native distributed APIs such as FSDP and TP, while integrating directly with Transformers, Diffusers, Accelerate, PEFT and TRL. The public beta will be available starting October 20, 2026, installable through the torch-neuronx package.

Hugging Face recommends that Neuron users work directly with its libraries instead of Optimum Neuron. The company says 89% of the 100 most-used text-generation model architectures have out-of-the-box support with TorchNeuron, compared with 46% in Optimum Neuron v0.4.5.

Sources: Hugging Face

Read the original story ↗

Cursor announced that cloud agents can now build interactive canvases as part of their work and share them with teams through a link. The company also made the Environments API available in beta, allowing users to programmatically create and manage the environments used by cloud agents.

Environments now support up to 300 repositories, while large environments load five times faster, according to Cursor.

Sources: Cursor

Read the original story ↗

Microsoft says it launched PRAISE, short for “Preceptorship for AI-accelerated Software Engineering,” an internal program that pairs early-career engineers with senior preceptors. Inspired by nursing’s preceptorship model, the program aims to help participants use AI tools thoughtfully and develop systems thinking, judgment, confidence and accountability for outcomes—not simply produce code faster.

According to Microsoft, the partnership with Carnegie Mellon University combines practical engineering experience with academic instruction to prepare the next generation. The company says computer science degrees remain relevant because understanding code, architecture and systems is still necessary even when AI writes part of the software; engineers must decide what deserves to be built and verify that it works safely and effectively.

Sources: Microsoft

Read the original story ↗

Figma announced Figma Specs, a Markdown-based editor for defining a brand’s visual language and bringing together design and code libraries, guidelines, canvas elements, and FigJam diagrams. The document can stand alone or exist within the canvas, serving as a reference for teams and first- or third-party agents to follow design rules when creating with AI.

The company also made code layers available in public beta, allowing users to add code to designs for interactive prototypes and ask the agent to turn designs into code or modify existing code. Figma said its MCP now supports more than 20 tools, including Open Code, Dia, Grok Bot, and Sarvam. Payload, the headless CMS acquired by the company last year, will be integrated with Figma Make to generate dynamic websites and manage content directly; Figma said Payload downloads increased sixfold over the past 12 months.

Sources: TechCrunch

Read the original story ↗

Hugging Face announced the availability of Black Forest Labs’ FLUX.3 Image Klein in Diffusers from launch day. The model supports text-to-image generation, image-to-image transformation with multiple references and layout control through bounding boxes described in the prompt.

The collection includes a distilled checkpoint for fast inference and a base model. According to Hugging Face, the distilled checkpoint generates 1,024 × 1,024 images in about two seconds on an H100, using 16 steps and no CFG. The package also includes NVFP4 and FP8 weights for smaller GPUs, LoRA training scripts for generation and editing, and a modular pipeline. The models can be tested in the Hugging Face space.

Sources: Hugging Face

Read the original story ↗

GitHub says its context-aware secret classifier improves detection quality with less noise. According to the company, the model increased weighted acceptance by 11.5% and reduced median scan execution time by 86%.

Sources: GitHub

Read the original story ↗

Google has begun allowing people aged 18 and older to submit Android apps made with Google AI Studio to a private Play Store library using a regular Google account, without creating or paying for a developer account. The apps do not appear in searches or recommendations and can be seen only by their creator; sharing with friends and family is planned but not yet active. The feature starts in the US and selected markets, and Google expects to add an option to move private apps to the public store in the coming months. The company says AI Studio includes safety checks, while the apps must still comply with existing Play Store policies.

Ask Play is expanding into general app and game search, with recommendations and comparison tables. Google also launched Game Finder in the US, using data from 1.2 billion players to create “Gamer DNA” profiles and suggest games; other markets will follow “soon”.

Sources: Ars Technica, Xataka

Read the original story ↗

Google announced a feature in the US that lets users launch personal Android apps created with AI Studio through vibe coding directly via Google Play. The change removes the need to connect a phone to a computer to install the apps; Google introduced the ability to create them in May.

The company says the apps will remain secure, private and separate from the store’s public catalog, and can be launched with a standard Google account. Google also plans to add sharing with a selected group of people soon. Sam Bright, Google Play’s vice president and general manager, said the feature could serve as an on-ramp for fledgling developers, who can move to a professional account with the same safety reviews and quality standards.

Sources: The Verge

Read the original story ↗

American Express introduced the Intent Intelligence Framework, designed to record and authenticate a customer’s instructions before an AI agent completes a purchase. The company says the technology captures constraints such as the merchant, maximum budget and preferences, then creates an Intent ID and Proof of Intent Token used to obtain a digital payment credential instead of the actual card number.

The framework has been added to the ACE Developer Toolkit, which American Express launched in April. The company told TechCrunch it has not yet seen an increase in disputes related to agentic commerce, but wants to anticipate the model’s growth. In an Amex Trendex report cited by the company, 66% of surveyed U.S. consumers said they were likely to use agents to research the best price, and 61% to receive recommendations, over the next 12 months. American Express hopes its approach will become an industry standard.

Sources: TechCrunch

Read the original story ↗

Cursor announced that the Origin App Marketplace now includes more than 20 apps, including integrations with Cloudflare, Linear, Sentry and Socket Security. The list covers deployments, continuous integration (CI), code review, security, monitoring and project management.

Starting now, any team building developer tools can create an app, integrate it with the Origin API and submit it for review by Cursor. Apps can receive webhook events and write check runs, pull request comments and reviews; after approval, they are listed in the marketplace.

Sources: Cursor

Read the original story ↗

Epic Games has reportedly shifted to an “AI-first” development model for an unannounced shooter being made with Disney, according to Bloomberg, as cited by Ars Technica and The Verge. Internal messages viewed by Bloomberg reportedly instructed the team to stop using hand-written code and produce the game’s art through fully AI-based workflows, including tools such as Anthropic’s Claude. Human developers would still manage and guide the code and visual assets generated by the models.

The game, reportedly featuring Marvel and Star Wars characters, was delayed indefinitely, according to Bloomberg; it had originally been scheduled for release in March 2027. An Epic spokesperson said the company is embracing AI-accelerated development, while CEO Tim Sweeney said Epic is moving developers around to expand the use of the technology. The multiyear partnership includes Disney’s $1.5 billion investment in Epic announced in 2024, according to The Verge.

Sources: Ars Technica, The Verge

Read the original story ↗

Gurmeet Saran, a member of Anthropic’s technical staff in San Francisco, is developing infrastructure to understand how Claude behaves and help ensure that its underlying architecture produces reliable outputs. In an interview with IEEE Spectrum, he said his work involves keeping data accurate, structured, and trustworthy while protecting private information from the start of the system’s architecture.

Saran says observing and understanding large language models remains a challenge: companies have decent tools for monitoring databases and web services, but are still developing comparable methods for these models. Before joining Anthropic, he worked on data quality and standardization at Yahoo and helped build analytics infrastructure at Intuit, experiences that inform his current work.

Sources: IEEE Spectrum

Read the original story ↗

Brandon Thomas, the creator of ArtCraft, launched WordCraft, GridCraft and DeckCraft as free, open-source alternatives to the Office suite. The apps were developed with help from Anthropic’s Claude and can be downloaded through GitHub for Windows, macOS, Linux and the browser. WordCraft edits documents and supports DOCX; GridCraft offers formulas, charts and XLSX support; and DeckCraft creates presentations but is still at an early development stage. The projects were developed in Rust and do not fully reproduce the features and formats of established suites.

According to Tom’s Hardware, Thomas wants to offer options for people who prefer not to depend on subscriptions. The programming community criticized the apps as “vibecoded,” a term used for projects produced with AI tools. On X, Thomas defended the technology as an aid and said he has 15 years of experience, arguing that AI models allow him to work “like an entire company.”

Sources: Tecnoblog

Read the original story ↗