Today’s lead · thehackernews.com
Compromised jscrambler 8.14.0 npm release drops a Rust infostealer
The jscrambler 8.14.0 npm package was pushed under a legitimate maintainer account with a preinstall hook that drops Windows, macOS, and Linux binaries. The payload targets cloud credentials, wallets, password vaults, browser/session data, and AI coding tool configs for Claude Desktop, Cursor, Windsurf, VS Code, and Zed.
Top signals
2 more
NVIDIA releases Nemotron Audex 30B-A3B, an open audio-text MoE model
NVIDIA introduced Nemotron-Labs-Audex-30B-A3B, a 30B MoE audio-text model with 3B active parameters for audio understanding, ASR, translation, text-to-speech, audio generation, and speech-to-speech generation. The linked model card says it supports thinking and instruct modes, up to 1M-token context, and inference recipes for vLLM, SGLang, and Transformers.
Meta removes Instagram AI image feature after public-content backlash
Meta pulled a Muse Image feature that let public Instagram content be referenced for AI-generated or altered images. The company said the creative tool "missed the mark" and is no longer available after user feedback.
Tools & repos
6 selected
SGLang v0.5.15
SGLang’s latest release focuses on production serving for GLM-5.2 NVFP4, with the project citing 500+ tok/s/user on 8x B300 and 450 on 4x GB300 at batch size 1. It also adds support for Hunyuan 3, HRM-Text, LocateAnything-3B, Unlimited-OCR, JoyEcho, and Qwen3.6, plus native web search, breakable CUDA Graph by default, MLA decode context parallelism, and FlashInfer all-to-all for routed MoE.
google-labs-code/stitch-skills
Google Labs’ TypeScript repo packages Agent Skills designed for the Stitch MCP server and compatible coding agents including Antigravity, Gemini CLI, Claude Code, and Cursor. It is useful for teams trying to standardize reusable agent behavior across editors and model providers.
davila7/claude-code-templates
A Python CLI for configuring and monitoring Claude Code. The high-signal use case is turning one-off Claude Code setup into repeatable project templates and giving teams a cleaner way to manage agent configuration.
AG KIT
AG KIT injects a .agents configuration folder into a project with 20 specialist personas, 45 skills, and 13 slash-command workflows. Its useful patterns are coordinator mode for multi-agent orchestration, persistent project memory, and context-compression routines that reduce repeated setup across coding-agent sessions.
JustVugg/colibri
Colibri is a zero-dependency C inference engine that runs GLM-5.2’s 744B-parameter MoE by keeping the dense core resident and streaming routed experts from disk. It is not fast, but it is a useful proof point for local MoE offload, SSD caching, and memory-management experiments.
Effects SDK
An SDK for adding AI video and audio effects to real-time apps. It is relevant for teams that need media enhancement features without building and serving the underlying effects pipeline themselves.
Blogs worth your time
3 reads
NVIDIA RoboLab: evaluating robot policies for real-world deployment
NVIDIA explains why robot policy benchmarks fail through visual overlap, fixed task sets, binary scoring, and weak statistics, then introduces RoboLab as a robot-agnostic simulation benchmark. The useful pieces are capability-tagged tasks, partial-credit scores, SPARC trajectory quality, failure event logging, and sensitivity analysis for diagnosing policy brittleness before real-world deployment.
Build a tool-backed market copilot with LangChain and Streamlit
The tutorial builds a market-brief generator that separates a LangChain/LangGraph agent engine from a Streamlit UI while grounding answers in EODHD data tools and OpenAI-generated summaries. It is a practical template for product teams that need flexible natural-language questions without letting the model invent financial facts.
A production-minded checklist for agentic coding setups
Jamon Holmgren lays out an opinionated agentic coding harness: AGENTS.md routing, self-healing docs, mandatory app runs, E2E tests, cross-model reviews, session worksheets, helper scripts, sweeps, visual regression, benchmarks, and end-of-shift validations. It is useful because it treats coding agents as a system of docs, tools, tests, and review loops rather than a prompt box.
Funding & acquisitions
2 moves
ideaForge
The listed Indian drone-tech company closed its QIP, allotting about 62.89 lakh shares at ₹795 per share. Proceeds are earmarked for working capital, debt repayment, product development, and general corporate purposes as ideaForge expands into precision-strike features and the Yeti cargo eVTOL platform.
Genrobotics
Zoho-backed Genrobotics reported FY26 operating revenue of ₹43.7 Cr, up 35% YoY, and PAT of ₹2.6 Cr. The robotics startup is seeking a Series B round mainly to expand manufacturing capacity for sanitation, medical, defence, and space robotics products.
Bengaluru radar
0 eventsThere are no relevant Bengaluru events to highlight today.