For people who build with AI.

According to the study described by Ars Technica, introducing coding agents increased the average time between a pull request’s submission and its merge by 49%. The share of pull requests receiving change requests nearly doubled, comments per request rose 35%, and the share of workers performing reviews increased 14%. The researchers could not attribute significant employment changes to AI.

Although 80% of the firms measured were using some form of AI code review by March 2026, agents accounted for just 23.3% of review comments and 10.8% of pull requests. The study indicates that humans still perform most reviews, creating a bottleneck that can offset the speed gains from generating code.

Sources: Ars Technica

Read the original story ↗

OpenAI announced new Codex capabilities, including reusable cloud development environments accessible from any device. The company says they can help tasks start faster and give teams a shared workspace with approved settings and permissions. The Codex command-line tool also gained voice controls and a view for delegating and tracking multiple tasks.

Codex is now integrated into a new code review experience in the ChatGPT desktop app, with summaries, change analysis and automatic reviews for GitHub pull requests and GitLab merge requests. OpenAI also announced Codex Security Cloud, which scans GitHub repositories, investigates findings and prepares fixes in the cloud, as well as API updates: the Decisions API, which uses Luna, and a version of the Agents API with support for computer use and Amazon Bedrock Managed Agents.

Sources: TechCrunch

Read the original story ↗

OpenAI announced plugin extensions that let developers build app-like experiences within ChatGPT. They can include a dedicated home in the sidebar, interactive panels for using tools while chatting, and file viewers compatible with the formats supported by each product. The company also introduced Plugin Creator, a new submission flow for the directory, and improvements to plugin recommendations and discovery.

Users will be able to approve each plugin’s requested permissions individually. OpenAI said plugins can also be hosted on ChatGPT Sites, allowing users to share them with colleagues who use their own connected data and permissions. The company also said it will support the proposed MCP Events specification, which could let plugins start automations based on events in connected apps, while making automations easier to add and manage.

Sources: TechCrunch

Read the original story ↗

An article published on Hugging Face presents ProvenanceGuard, a post-generation verification layer for AI agents based on the Model Context Protocol (MCP). The system aims to detect “cross-source conflation”: cases in which a claim is supported by the combined evidence but attributed in the answer to the wrong source, a problem conventional verifiers may miss when they analyze all evidence together.

ProvenanceGuard processes MCP traces, preserves source identifiers, breaks the answer into claims, checks which source supports each one, and compares that source with the one named or implied by the answer. It then produces per-claim results and an overall decision to allow or block the text; blocked answers can go through repair and another verification pass. In the experiments described in the article, local models were used to find relevant sources, check textual support, and decompose answers, with additional checks for numbers, dates, and identifiers.

Sources: Hugging Face

Read the original story ↗

Cloudflare is developing CryptoLabe, an internal AI-powered tool to discover where cryptography is used across its codebase, identify dependencies and track progress toward post-quantum migration. The company says it aims to achieve full post-quantum readiness by 2029; many products already use post-quantum encryption over TLS 1.3, but work remains on authentication and on other connections and uses of public-key encryption.

Cloudflare says CryptoLabe should also help teams understand how to upgrade systems, measure classical and post-quantum cryptography use by repository and product, and identify protocols, libraries or standards that lack a defined migration path. The tool is specialized for the company’s internal systems, is still evolving and will not be made available to customers.

Sources: Cloudflare

Read the original story ↗

Shopify announced that browser-based AI agents can now complete purchases on merchants’ sites, expanding its previous WebMCP support, which allowed agents to search for products and add them to carts. The company said the checkout integration, including Shop Pay, lets agents read and update the checkout screen and submit the transaction with the buyer’s authorization, without relying on screenshots or webpage scraping.

The update adds the get_checkout, update_checkout, and complete_checkout tools, allowing agents to inspect checkout, change details such as the customer’s address and delivery option, and place an order after authorization. The feature is rolling out to all eligible merchants. Shopify also operates an MCP server for server-to-server operations; WebMCP is designed for agents working in the browser and uses the Universal Commerce Protocol (UCP). TechCrunch reports that Shopify also announced a direct partnership with Instinct for agentic commerce.

Sources: TechCrunch

Read the original story ↗

Cloudflare updated Kitesurf, its browser for AI agents that runs entirely on Cloudflare Workers, with WebMCP support, improved DOM performance and terminal-based rendering. WebMCP lets websites expose functions directly to agents, which can use them instead of simulating clicks; the feature can be tested in Kitesurf’s public playground with Cloudflare Radar.

The company also added support for APIs and standards including CSS Layout, CSSOM, CSS Typed OM, custom elements, JSON modules, import maps and URL-based module loading. Cloudflare says Kitesurf now passes more than 730,000 Web Platform Tests subtests, 500,000 more than at launch, while reducing the work involved in DOM traversal, timers and script loading.

Sources: Cloudflare

Read the original story ↗

Nvidia launched the Open Agent Safety Platform, combining OpenShell, open-source software that defines what each agent can access, with Sentry, an independent monitoring layer running on BlueField-4 processors. Nvidia says the system checks restrictions before and during tasks and can quarantine agents that attempt to cross their boundaries within milliseconds. OpenShell, announced in March, is now generally available.

The company says more than 100 companies support the effort, including Anthropic, Microsoft, and SpaceX. OpenAI is not among the public supporters, but a spokesperson told TechCrunch that the company is working with Nvidia on agent security and OpenShell. Nvidia executives say the platform could have prevented OpenAI agents from breaching Hugging Face; that is the company’s assessment, not a demonstrated result. Nvidia is also working with Arm and Intel to adapt the system to other architectures, although Sentry depends on proprietary Nvidia hardware.

Sources: Wired, g1, TechCrunch, Tecnoblog, The Verge

Read the original story ↗

Cybersecurity firm UpGuard found about 16,000 databases hosted by Supabase that exposed some personal data, the company told TechCrunch. The publicly accessible information included names, addresses, phone numbers, passwords and some authentication tokens; the data covered private conversations, vehicle license plates and contacts from immigration services. UpGuard also identified a database linked to an African government consulate in France and another used to intercept text messages from a virtual SIM operation. Most of the datasets appeared to be in the United States, but the firm said the problem is worldwide.

The episode highlights risks in applications built with artificial intelligence, whose generated code can contain flaws or require configurations developers do not understand. Supabase Chief Information Security Officer Bil Harmer told TechCrunch that projects are secure by default and that security is a shared responsibility with customers, who control their projects’ configuration.

Sources: TechCrunch

Read the original story ↗

PrismML created a 2-billion-parameter version of its 1-bit Bonsai language model for smart glasses based on Qualcomm’s Snapdragon AR1 Gen 1 Platform. Qualcomm demonstrated the model at its Snapdragon Summit; it can run locally and let users ask in real time what they are looking at. PrismML says its technology shrinks larger models by four times while retaining almost all of their performance on standard benchmarks. The startup says it is pursuing open-weight AI that runs on devices, but no smart glasses using the technology have been announced so far.

Sources: TechCrunch

Read the original story ↗

Liquid AI released LFM2.5-VL-DSpark, an experimental drafter model for its LFM2.5-VL-3B vision-language model. The company says speculative decoding speeds generation without changing output quality, with gains of up to 3.13x on-device and 2.66x on an H100, while end-to-end latency improves by up to 2.62x and 2.27x, respectively. The drafter adds about 280 million parameters, or 8.9% to the 3-billion-parameter model.

The release includes day-one support for llama.cpp, MLX-VLM and SGLang. Liquid AI evaluated six vision tasks, including image and chart question answering, captioning and complex reasoning. The technique accelerates decoding only; image encoding and prefilling remain unchanged, which can limit the overall gain when those stages account for a large share of response time.

Sources: Hugging Face

Read the original story ↗

According to Xataka, software engineer @v0xium says that at a large company whose name he did not disclose, Claude Code produces specifications, code, tests, documents, tickets and reports. Employees spend their days validating the results while facing pressure to deliver more, with some working 12- to 13-hour days. In an interview with Business Insider, he said the company measures performance through development cycles, pull requests and shipped features—metrics that do not always indicate product improvements.

Boris Cherny, who leads Claude Code, says AI-generated code still requires rigorous controls and reviews. Xataka also cites a METR study that found AI tools increased coding time by 19% for certain tasks performed by experienced open-source developers in the first half of 2025. The engineer says the lack of time to understand code and investigate problems makes the work more stressful and less fulfilling.

Sources: Xataka

Read the original story ↗

Hugging Face presents a workflow for moving an SO-101 arm from a conventional MuJoCo setup to as many as 2,048 parallel environments with MuJoCo Warp (MJWarp). The technology uses NVIDIA Warp to implement MuJoCo physics and compile CUDA kernels, allowing many worlds to advance simultaneously on the GPU while keeping simulation data close to the device.

The article also explains that NVIDIA Warp lets developers write kernels in Python, with CPU or CUDA execution, autodifferentiation, and PyTorch and JAX interoperability. The material focuses on preparing and scaling the simulation environment; it does not train a policy. Later integrations with Newton and Isaac Lab are described as the next layers of the workflow.

Sources: Hugging Face

Read the original story ↗