Today’s lead · thehackernews.com
OpenClaw flaws enabled a WhatsApp-to-host AI assistant attack chain
Researchers detailed three now-patched high-severity flaws in the OpenClaw personal AI assistant, including command injection and a sandbox path traversal issue that could enable credential theft, privilege escalation, and arbitrary code execution on the host.
Top signals
4 more
Injective Labs GitHub compromise pushed wallet-stealing npm packages
Threat actors compromised the Injective Labs SDK GitHub repository and published @injectivelabs/sdk-ts@1.20.21 with fake telemetry that exfiltrated wallet private keys and seed phrases; 17 additional scoped packages pinned the malicious SDK version.
Six U-Boot flaws can crash devices or run code before signature verification
Binarly found six flaws in U-Boot affecting FIT image parsing: four can crash devices and two could allow code execution before the bootloader verifies the software signature. The vulnerable code has been present since v2013.07 across more than 50 stable releases and many vendor firmwares.
OpenAI is changing ChatGPT Work and Codex defaults after launch regressions
OpenAI said it reset usage limits across ChatGPT Work and Codex, changed defaults and the model picker to avoid steering users toward unnecessarily expensive settings, fixed plugin submission issues, and will restore chats and projects to the sidebar next week.
SK Hynix raises $26.5B in a record U.S. foreign IPO amid AI memory demand
SK Hynix raised $26.5 billion in its U.S. market debut, described as the largest-ever U.S. debut by a non-American company, as AI chip demand keeps investor appetite high and U.S. pressure grows for Korean memory makers to build local fabs.
Tools & repos
7 selected
Sim
An open-source workspace for AI agents and workflows. It is worth a look for teams that want a visual, inspectable layer for composing agent workflows instead of burying orchestration in ad hoc scripts.
wonderwhy-er/DesktopCommanderMCP
An MCP server for Claude that gives it terminal control, file-system search, and diff file editing. Useful for local desktop-agent workflows, but the shell and filesystem permissions deserve the same scrutiny as any privileged automation tool.
langchain-ai/open-swe
OpenSWE is an open-source asynchronous coding agent. The post says it is model-agnostic, fully OSS, and integrates with LangSmith for observability, making it a useful reference for internal coding-agent systems.
Claude Code in-app browser
Claude Code desktop now has a sandboxed in-app browser that can open docs, designs, and sites, then read, click, and interact with them like local dev servers. Sessions can be configured for persistence.
ax
A curl-like CLI for coding agents to fetch, discover, and extract from web pages with structured reports and CSS-selector rows. The practical win is avoiding raw HTML dumps and throwaway Python inside agent sessions.
Unsloth Qwen3.6 NVFP4 quants
Unsloth released Qwen3.6 NVFP4 quantizations claiming 2.5x faster GPU runs. The post says Qwen3.6-27B fits 24GB VRAM and highlights improvements for tool calling, agent use, and looping.
vLLM hardware-aware Dynamic SD
Cohere’s hardware-aware dynamic speculative decoding has been merged into vLLM. It adapts draft-token behavior to batch size and hardware, with fallback when speculative decoding would otherwise slow inference down.
Blogs worth your time
7 reads
Reducing High-Bandwidth Memory Bottlenecks in JAX-Based LLM Training with Host Offloading
NVIDIA explains how JAX host offloading moves selected activations to pinned host memory and streams them back to relieve HBM pressure. The DeepSeek-V3 671B MaxText run on GB200 NVL72 shows why memory placement and overlap are now first-class training knobs.
Better tools made Copilot code review worse. Here’s how we actually improved it.
GitHub shows that swapping Copilot code review to shared grep, glob, and view tools initially raised cost and missed more issues until the workflow was rewritten around pull-request evidence. The useful lesson: agent tools do not help unless the operating procedure matches the task.
OpenWiki Brains: Proactive Memory for AI Agents
LangChain introduces OpenWiki Brains, a local Markdown wiki memory layer that ingests sources like Gmail, Notion, Git, X, Hacker News, and web search. It is a concrete pattern for giving agents refreshed work context without relying only on reactive chat memory.
Kernel Fusion in NVIDIA CUDA: Optimizing Memory Traffic and Launch Overhead
This walks through sum(abs(x)) from a naive two-kernel implementation to manual fusion, torch.compile, and cuda.compute. It is a practical reminder that avoiding global-memory round trips can matter as much as choosing a faster model kernel.
Profiling in PyTorch (Part 3): Attention is all you profile
Hugging Face uses naive attention, in-place operations, PyTorch SDPA, and kernelized attention to teach how attention optimizations show up in profiler traces. Useful for engineers who need to read GPU timelines instead of blindly toggling backends.
Behavioral Privacy Leakage in Agentic Negotiation
Apple formalizes how negotiation agents can leak private constraints through observable behavior such as concession trajectories and timing, even when explicit values are protected. The proposed randomized policy cut adversarial inference accuracy by 43–50% in synthetic negotiations while keeping utility and success rates above 90%.
GPT 5.6 Has 72 Possible Configurations. What's A Good Default?
Sebastian Raschka maps GPT-5.6’s Work/Codex, Sol/Terra/Luna, six effort levels, and Standard/Fast choices into 72 configurations. It is a concise framing for teams setting cost-aware defaults instead of reflexively picking the highest tier.
Funding & acquisitions
5 moves
Zetwerk
The Bengaluru manufacturing platform received SEBI approval for its IPO, expected to include a fresh issue and an offer for sale. Zetwerk connects industrial customers with supplier and manufacturing networks and uses Zetwerk OS for sourcing, production planning, supplier coordination, and project execution.
Indifi
The Gurugram SME lending platform is set to raise ₹79 Cr, around $8.3 Mn, in its first funding round in three years. The round values the company at ₹1,418 Cr and proceeds are planned for expansion, infrastructure, capex, working capital, debt repayment, and general corporate purposes.
Iswarya Fertility Center
The Chennai-headquartered fertility network raised over ₹350 Cr, about $37 Mn, in growth funding. The company plans to add centres in existing and underserved markets, strengthen clinical infrastructure, invest in digital capabilities, and expand its embryology talent pipeline.
Credit Fair / BlackSoil
BlackSoil acquired Credit Fair’s rooftop solar financing business in a ₹45 Cr cash slump sale. The deal gives BlackSoil a granular retail lending book and entry into rooftop solar financing after Credit Fair struggled to raise capital amid tighter RBI rules.
BiofuelCircle
The Pune biomass and biofuel marketplace raised ₹35 Cr in fresh funding. BiofuelCircle connects farmers, suppliers, and industries with price discovery, logistics, and trade finance for agri-residue and biofuel transactions.
Bengaluru radar
3 eventsDataHack Summit 2026
Analytics Vidhya’s Bengaluru AI conference returns with the theme “Human × AI: The Rise of the Agentic Operating Layer.” The event is positioned around production AI, agentic systems, and practical industry adoption rather than generic AI awareness.
WeAreDevelopers Conference India
A Bengaluru developer conference for engineers, engineering leaders, and teams building production AI systems. Useful for teams looking beyond demos into AI-assisted development and software engineering practice.
Great International AI Native Summit (GAINS)
An engineering-first AI conference in Bengaluru covering agents, AI-assisted development, infrastructure, reliability, security, governance, and production operations. The positioning is squarely for builders shipping intelligent systems.