For people who build with AI.

According to an Anthropic guide, Fable 5.2 tends to behave differently when it believes the code being edited is its own. In that situation, the model often merges repeated sections and makes related fixes without specific instructions; when it does not know who owns the code, it may only add duplicate code and flag problems instead of fixing them.

Anthropic tested the guidance on small programming tasks. With the phrase “This code is yours,” Fable 5.2 merged repeated code in about half of attempts, versus fewer than one in ten without the instruction. The company cautions against using the phrase when the goal is to change as little code as possible, since the model may make changes that were not requested.

Sources: Xataka

Read the original story ↗

Security researchers at Apiiro found more than 100,000 Claude Code “.claude/settings.local.json” files uploaded to public GitHub repositories. About 1,100 files from 63 organizations contained secrets such as database passwords, API tokens, private keys, SSH server logins and AWS keys. The issue occurs when users choose “Yes, and don’t ask again”: the approved command can be saved literally, including credentials. Although the file is normally excluded by the developer’s global “.gitignore”, that protection may not apply at the repository level, and Claude Code can commit the file without user input, Apiiro says. Secrets can also persist in repository forks after the original file is removed.

According to Xataka, Boris Cherny, who leads Claude Code, acknowledged on Hacker News that the tool did not scrub secrets from saved content and recommended using automatic mode and enabling repository secret scanning. He said the team is testing automatic-mode leak detection and a basic built-in scanner, with no announced timeline.

Sources: Ars Technica, Xataka

Read the original story ↗

When announcing the second release candidate for Linux 7.4, Linus Torvalds said that, at least in the eBPF area, fixes may be arriving faster than new problem reports, citing Alexei Starovoitov, one of the technology’s maintainers. Torvalds added that this would be positive “if true.”

eBPF allows software to run in a controlled way inside the kernel. Torvalds considers using AI to find bugs the “new normal” in Linux, but the excess reports—including low-priority issues or problems that have already been fixed—could still overwhelm the team. Linux 7.4 is expected in December 2026.

Sources: Tecnoblog

Read the original story ↗

Automattic, the owner of WordPress.com, launched Spacefast, a platform for deploying websites, apps and services built with vibe coding. The service provides infrastructure including databases, authentication, file uploads, email and analytics, and can connect to agents and chatbots such as ChatGPT, Claude, Grok Bot and Cursor, or to a Git repository. The platform competes with services such as Lovable and Emergent, but lets users choose which AI tools they use. The product is led by Batuhan İçöz, who initially developed it during an internal company project; Automattic soft-launched it in September to gather feedback.

According to İçöz, people and organizations have already used Spacefast for internal mockups, presentations and offer-related websites. Automattic also plans to draw on existing products such as WordPress, web analytics and WooCommerce to serve enterprises seeking to reduce tool sprawl. The executive acknowledged that some users may leave after trying the service, and said the priority is to make building simple and the results better than they would be without the platform.

Sources: TechCrunch

Read the original story ↗

In a post published on the GitHub Blog, GitHub says it saw more than 1.2 billion commits in one week in October, over 400% more than a year earlier. The post cites a Faros AI study that found pull requests spent 441% longer in review when teams moved from low to high AI adoption, arguing that faster code generation does not remove the work of planning, reviewing and monitoring production.

The guide proposes using agents across the Plan, Build, Verify and Observe stages. Examples include bringing context together from connected tools, collaborating with Copilot in Slack or Teams, turning requirements into GitHub Issues, triaging tasks and comparing approaches in Plan Mode. It also describes session continuity across devices and task execution in cloud environments, while emphasizing that developers still need to approve plans, evaluate results and make complex decisions.

Sources: GitHub

Read the original story ↗

Vermeij, a former Rockstar technical director, told Rock Paper Shotgun that he would be concerned someone could take the PS5 version of GTA VI and use AI to adapt it for PC. He said reverse-engineering it would probably cost tens of thousands of dollars in tokens and produce a port inferior to Rockstar’s work, but potentially good enough for some PC players.

According to Ars Technica, AI reverse-engineering tools may quickly turn a console game binary into code that is easier to adapt, weakening the traditional advantage of a later PC release. There is no official timing announcement for GTA VI’s PC port; Ars Technica says reports suggest Rockstar is aiming for the first quarter of 2027.

Sources: Ars Technica

Read the original story ↗

Cursor now offers real-time voice calls with agents in the Agents Window, the IDE, on the web, and in its iOS app. Users can describe changes, ask about work in progress, and receive updates while an agent performs coding tasks; on iOS, they can also start a call by saying “Hey Siri, call Cursor” and control audio from the Lock Screen.

The voice call feature is powered by Grok Voice and lets users choose voices including Ara, Eve, Leo, Rex, and Sal. Cursor says it runs with zero data retention when Privacy Mode is enabled. The feature is separate from dictation, which continues to convert speech into text in the chat input.

Sources: Cursor

Read the original story ↗

Microsoft says advances in AI are accelerating the discovery, analysis and remediation of Windows vulnerabilities. The company uses MDASH, an agentic code scanner that combines multiple AI models and specialized agents to find and validate potential flaws; security and engineering teams review the findings and determine the response.

According to Microsoft, AI also helps assess impact, identify affected Windows versions and code variants, choose remediation strategies, and build security and regression tests. The company says it applies similar capabilities to validation-test selection, creation, triage and analysis, while keeping human review and engineers’ responsibility for final decisions and outputs.

Sources: Microsoft

Read the original story ↗

According to The Verge, Anthropic changed Claude Code’s internal instructions to allow original games, games in a genre, or games based on designs and assets supplied by the user, while preventing the reconstruction of a specific game—even under another name or presented as original. The company confirmed the change. Addy Osmani, from Anthropic’s technical team, said two sentences were added to the Claude Code prompt concerning Opus 5.5; users can contact support if the tool wrongly rejects an allowed request.

The measure responds to concerns about Opus 5.5’s ability to reverse-engineer code, which PC Gamer highlighted. According to Xataka, developers canceled public demos for games including “Drunken Goddess Reflux” and “ReVamp” out of fear of AI-generated copies appearing before launch; “Needle in a Haystack Simulator” was reportedly targeted by numerous clones.

Sources: Xataka

Read the original story ↗

Hugging Face says a bug introduced in Transformers on July 30, 2026, allowed UMT5’s decoder to attend to future tokens during training. The problem remained in the main branch for almost two months, but a small UMT5 model running on a CPU runner identified the regression when the commits were replayed; the fix came on September 28. The reduced model used 66 MB, compared with 1,227 MB for the source checkpoint.

The company is proposing Model Integration Contracts (MICs), a standard with three parts: a reduced model derived from the real checkpoint, a runnable usage path and human-reviewed recorded baselines. Published in huggingface-mic and pinned to frozen revisions, the contracts can be selected based on the files changed in a pull request. Hugging Face says they complement, rather than replace, full integration tests.

Sources: Hugging Face

Read the original story ↗

Microsoft says nine engineers merged 1,211 pull requests into the Copilot Cowork backend’s main branch on October 21. The company says it took the product from an empty repository to worldwide general availability in about four and a half months, with agents writing most of the code. Copilot Cowork reached its first enterprise customers through the Frontier program on March 30 and general availability on June 16, according to Microsoft.

The team began requiring a Markdown specification before each change, covering the task’s purpose, how success would be measured and what could not break. Agents handled narrow tasks within defined boundaries, while engineers reviewed higher-risk changes and retained responsibility for decisions. Microsoft also says it moved the live runtime from Python to Rust between September 23 and October 10 without interrupting the service.

Sources: Microsoft

Read the original story ↗

Anthropic updated Claude Code’s system prompt on Opus 5.5 to steer the tool away from recreating specific games made by others, saying that using Claude to infringe intellectual property violates its usage policy. Company engineer Addy Osmani said original games, games in a genre, and projects based on users’ own designs and assets should remain allowed. Ars Technica reports that the rule also appears in Claude Code version 2.1.314, based on independent verification by XDA, although reports on Anthropic’s GitHub suggest it has interfered with some legitimate uses.\n\nThe change comes amid a wave of AI-assisted decompilation and porting projects involving older games. According to Ars Technica, the increase in these projects contributed to the shutdown of decomp.dev, a community dedicated to decompilation whose maintainer said the community aspect of the work had been undermined.

Sources: Ars Technica, The Verge

Read the original story ↗

According to Xataka, videos show iOS 16 booting on Android phones, including a Redmi Note 12. The article's technical explanation details an example using the Redmi Note 10 Pro: the open-source ios-sweet project replaces the Linux kernel with Apple's XNU and uses a hardware description copied from an iPhone X to make the system believe it is running on an Apple device. The article's headline says Claude helped, but the text does not explain that role.

On the Redmi Note 10 Pro, the lock screen, home screen, Control Center, Settings, touch input and physical buttons work. Wi-Fi, Bluetooth, mobile data, audio, cameras and biometric sensors do not; some apps show a black screen. Without a graphics driver, the interface runs at about 20 to 30 frames per second, and booting takes four to five minutes, according to Decrypt. Android Authority cites versions adapted for the Redmi Note 12 Pro and LG V60, as well as work on other models.

Sources: Xataka

Read the original story ↗

Xataka says vibecoding is making it possible to create tailored applications through a conversational interface with a language model, without requiring users to learn programming. The article cites examples such as apps for controlling smart-home devices while keeping data on a local network, removing bloatware from Android phones and managing family tasks with a database on the user’s own server. The software can also be adapted so data stays in the user’s environment.

According to Xataka, this does not eliminate control over applications: it shifts it from developers to AI models and companies such as OpenAI, Anthropic, Google and xAI. Subscriptions may continue to exist, now tied to an AI plan, while support, security, maintenance and bug-fixing limitations remain. The site says vibecoding democratizes app creation but does not replace developers who can build robust products for thousands of users.

Sources: Xataka

Read the original story ↗

NVIDIA presents projects in which developers guide AI agents with natural-language instructions to assemble assets, connect physics and rendering, and check how scenes behave in Omniverse. The company says its libraries provide GPU-accelerated physics, rendering and sensor-simulation capabilities, while the agents help turn ideas into applications for exploring scenarios, investigating failures and evaluating designs.

The examples include reconstructing buildings from panoramic photos and point clouds, a what-if analysis application for a gas turbine and generator, and creating a kitchen scene with SimReady assets. NVIDIA says the SimReady Central Library provides access to more than 2,000 simulation-ready USD assets. The projects use models such as GPT-6 Astra and tools including OpenUSD, Cantera and SciPy.

Sources: NVIDIA

Read the original story ↗

Google announced the Connected Apps Platform, which lets businesses and developers build, test and submit connected apps for use in Gemini, Google Search or both. The public preview is available in beta to U.S. developers who join the waitlist. To participate, they must integrate a cloud-hosted MCP server or Android MCP using the AppFunctions APIs and register the app in Google Connected Console.

After publication, developers can track traffic, latency and error rates, while maintaining control over elements of their brand identity. Google says more than 50 connected apps are already available in Gemini, including services from Adobe, Experian, Localiza and Peloton; they will come to Google Search over the next few weeks. The Calm app is also expected to launch on Gemini and Search. Google did not say when the platform would become more broadly available.

Sources: TechCrunch

Read the original story ↗

Thomas Dohmke, who was GitHub’s CEO from 2021 to 2025, has officially launched Entire, a Git hosting and collaboration platform built to compete with his former company. The service was introduced in July and is designed for human developers and artificial intelligence agents to work together. Instead of pull requests, the platform uses Trails, which record what changed in the code and why.

Entire says its Git-compatible version-control server enables pushes up to 89 times faster than GitHub’s and six times faster than Cursor Origin’s. The platform does not charge for repositories or Trails, but requires credits for AI agents; each account receives US$100 in welcome credits.

Sources: Tecnoblog

Read the original story ↗

Microsoft says it is developing and running AI agents to transform large codebases, including Windows and Azure. The work it cites includes migrating a 400-million-line codebase to a new compiler toolchain and modernizing legacy C and C++ code into Rust, with increasing use of formal verification.

According to the company, generating code is no longer the hardest part: teams must verify that changes preserve behavior, security and architectural intent. Since 2024, Microsoft says it has combined agents with static analysis, testing and formal methods; by May 2025, it says agents had become better planners than its workflows. The company also says human attention has become the scarcest resource in agentic software engineering.

Sources: Microsoft

Read the original story ↗