3 Italian passengers fly out of Delhi without immigration clearance: What exactly happened

Gurugram Hit-and-Run Lawyer Blames Biker, Claims Car Was Being Followed

A lawyer involved in a hit-and-run incident in Gurugram has shifted blame onto the biker who was struck, alleging the car was being followed before the crash. The biker, a woman who goes by 'Rebel Wheels Sia' on Instagram, is an active member of a biking group. She claims the car had been trailing her group prior to the collision. The lawyer suggested the biker sought the confrontation to gain social media followers. The incident has drawn significant public attention amid conflicting accounts from both sides.

Amazon Research Questions Reliability of LLM-Based Evaluation Judges

Amazon Science has published research examining whether agreement among large language model (LLM) judges can be trusted as a reliable evaluation signal. The study investigates a growing practice in AI development where LLMs are used to assess the quality of other models' outputs. Researchers raise concerns that consensus among LLM judges does not necessarily indicate correctness or accuracy. The work highlights potential blind spots and shared biases that could lead multiple LLM judges to agree on flawed assessments. The findings have implications for how the AI community designs and interprets automated evaluation benchmarks.

Cheap Solar Panels Are Reshaping the Global Energy Landscape

Solar power is experiencing a dramatic cost decline, driven largely by the mass production of Chinese-manufactured panels. These affordable panels are being installed on rooftops worldwide, altering the economics of energy generation. The shift is putting significant pressure on large traditional utility companies, which are being forced to adapt their business models. The widespread adoption of solar is changing fundamental assumptions about how electricity is produced and distributed globally.

Twigg

Discussion | Link

Developer Builds Open-Source RAG Platform to Unlock Information Trapped in Documents

A developer with years of experience delivering document-heavy digital services has released AI-DocumentIntelligence, an open-source, self-hostable platform for natural-language question-and-answer queries over uploaded documents. The system ingests PDF, DOCX, and TXT files, splits them into chunks, stores embeddings in PostgreSQL using pgvector, and returns answers citing source passages. Built on React, Node.js, and LangChain, it supports both OpenAI and Anthropic Claude as interchangeable LLM providers controlled by a single environment variable. The project addresses a common bottleneck in public sector and regulated enterprises, where critical information is technically available in documents but practically unsearchable. The developer prioritised provider-agnosticism and auditability over novelty, using a deliberately familiar tech stack to lower the barrier to adoption for existing teams.

Microsoft releases AI code of conduct banning hacking and deception by its models

Microsoft has introduced a formal code of conduct outlining behavioral standards that its AI models are expected to follow. The guidelines include broad principles such as supporting and augmenting human capabilities rather than replacing them, and promoting human well-being. The code also establishes specific safety constraints designed to put those principles into practice. Among the key prohibitions are actions like hacking systems or deceiving users. The move reflects growing industry pressure to define clear ethical boundaries for AI systems.

AI Industry Pushes for Controlled Development as Safety Frameworks Fall Behind Capabilities

A growing coalition of AI lab leaders, security researchers, and policy advisors is calling for a deliberate slowdown in frontier AI development, arguing that safety evaluation and compliance frameworks cannot keep pace with rapid capability growth. Anthropic's Dario Amodei and other industry figures warn that as models begin contributing to their own training pipelines, the window for meaningful safety validation narrows sharply. The proposed approach does not call for halting research entirely, but instead introduces mandatory operational checkpoints such as independent pre-deployment audits, granular system access limits, and robust activity logging before models are released. From a systems engineering perspective, connecting autonomous AI agents to core business infrastructure — including databases, payment systems, and customer records — significantly expands cybersecurity attack surfaces overnight. Proponents compare this regulatory model to established safety standards in aerospace and pharmaceuticals, framing such controls as fundamental operational hygiene rather than an obstacle to innovation.

WebMCP Brings MCP Tool Registration to the Browser Without a Working Directory

A developer has published Part IV of a series exploring how AI agents can access context when there is no working directory, file system, or server process — only a browser tab. The approach uses WebMCP, an experimental Chrome API, alongside a polyfill package to register tools directly from a web page via document.modelContext.registerTool(). The demo site faf.one/webmcp exposes three tools: a form-based context collector, a client-side WASM scoring engine, and a minimal AGENTS.md renderer. Security is enforced through a strict URL allowlist limited to two hosts, HTTPS-only fetches, and outright rejection of any URL containing an MCP endpoint path. The project deliberately distinguishes itself from the full faf-cli compiler, noting that the browser-based renderer omits advanced features like guardrails tiers and branch-aware commit instructions.

Developer stress-tests Meta Muse AI agent until control plane times out

A developer conducted black-box stress testing on Meta Muse, an AI agent platform by Meta. The testing pushed the system to its limits, eventually causing the agent control plane to experience timeout failures. The findings were documented and published on the developer's personal blog. The experiment highlights potential reliability and scalability concerns with Meta Muse under heavy load conditions.

Researchers Expose Security Vulnerabilities in AI-Powered Customer Service Agents

Security researchers at Intigriti have published findings on vulnerabilities present in AI-driven customer service agents. The research highlights how these systems can be manipulated or exploited by malicious actors. Such weaknesses pose risks to businesses and customers who rely on AI agents for support interactions. The blog post, shared on Hacker News, outlines hacking techniques and tools relevant to testing these AI systems. The findings underscore the growing need for robust security practices as AI customer service adoption expands.

Dangote Opens Africa's Largest Share Sale to Ordinary Investors

Nigerian billionaire Aliko Dangote, Africa's richest man, has launched the continent's biggest share sale. The offering is tied to his oil refinery, a major industrial project in Nigeria. Dangote has stated his intention to make the opportunity accessible to everyday investors, not just institutions. The move aims to allow ordinary people to benefit financially from the refinery's success.

How my e-reader lost its stripes

Article URL: https://www.serpentine.com/posts/2026/x3-stripes/ Comments URL: https://news.ycombinator.com/item?id=49699489 Points: 16 # Comments: 1

Developer Builds Open-Source AWS DevOps Agent to Help SRE On-Call Engineers

A developer has launched an open-source project called SentinelOps, an AWS-focused DevOps agent designed to assist site reliability engineers during on-call incidents by generating LLM-based root-cause hypotheses. The project is in its early stages, with the creator marking this as day one of a 100-day development journey. The tool is built for Kubernetes environments and can be deployed via a single Helm install command. The developer is actively seeking feedback from on-call engineers, either through brief conversations about real incidents or by testing the tool directly. The project is publicly available on GitHub at github.com/maxuver/sentinelops.

How to Build a Production-Ready Invoice Extraction Pipeline with Azure AI Tools

A technical guide published on DEV Community outlines how to build a robust invoice extraction pipeline using Azure AI Document Intelligence and Power Automate, going beyond basic API demos. The pipeline covers five stages: document ingestion, classification, extraction, confidence-based routing, and posting to a system of record or review queue. The author highlights that many tutorials rely on the now-deprecated Analyze Invoice action, recommending the current v4.x Analyze Document connector instead. A key focus is confidence-based routing, where documents are directed to automated processing, human review, or rejection based on model confidence scores and business rules. The guide also warns that sending non-invoice documents like statements or remittance advices to an invoice model can produce misleadingly confident but incorrect results.

Mukesh Kumar joins Nottinghamshire for two County Championship fixtures

India and Bengal fast bowler Mukesh Kumar has signed with Nottinghamshire for two County Championship matches. The pacer will feature in the side's games against Sussex and Yorkshire. The short-term county deal offers Mukesh valuable red-ball experience in English conditions. Such stints are increasingly common among Indian cricketers looking to sharpen their skills abroad.

MCP Registry Has 31,309 Unique Servers, Not 101,219 — 23% Lack Source Code

A developer analysed the official Model Context Protocol registry via its public API on 14 September 2026, finding that the widely cited figure of over 100,000 MCP servers is inflated by roughly 3.2x due to historical versioning. After deduplication, the registry contains just 31,309 distinct servers, with 69% of entries representing superseded older versions. Nearly 23% of servers — about 7,090 — have no linked source repository, making them impossible to audit without directly connecting to them. Concentration is also notable, with three hosts alone accounting for 14.5% of all servers and 17.4% sharing a hostname with at least 50 others. Monthly new server registrations surged from 456 in September 2025 to 6,673 by August 2026, raising concerns about the scalability of any manual review process.

Hyderabad's Lee-Chee Ka Raja idol dazzles with 24K gold and 3.8 lakh diamonds

A jewel-encrusted Ganesh idol is drawing large crowds in Hyderabad during Ganeshotsav 2026. The Lee-Chee Ka Raja features 24-karat gold plating and approximately 3.8 lakh American diamonds. Its design draws inspiration from Mumbai's iconic Lalbaugcha Raja and the famous Tirupati Balaji temple. The idol is located in the historic Begum Bazar area, allowing visitors to explore nearby heritage sites. Authorities are advising festival-goers to use public transport when travelling through the old city.

MIT creates method to force AI to comply with safety rules

Article URL: https://theframenews.org/en/hardflow-safety-critical-ai/ Comments URL: https://news.ycombinator.com/item?id=49699434 Points: 5 # Comments: 0

Trump administration moves to repeal Biden-era emission limits on power plants

The Trump administration is planning to scrap emission restrictions on coal and gas power plants that were introduced during the Biden era. The Environmental Protection Agency is expected to formally announce the final rule in the near future. These regulations had targeted one of the largest sources of greenhouse gas emissions in the United States. Environmental organizations have voiced strong opposition to the rollback, raising concerns over public health and safety. Legal challenges from advocacy groups are widely anticipated following the policy reversal.