Today’s lead · x.com
Anthropic reportedly tightens Claude access against Chinese-company workarounds
Cointelegraph, citing the Financial Times, says Anthropic is moving to close loopholes that let Chinese companies access Claude through workarounds.
Top signals
5 more
North Korea-linked npm packages mimic Rollup tooling to steal developer secrets
JFrog linked malicious npm packages such as rollup-packages-polyfill-core and rollup-runtime-polyfill-core to a campaign that used lookalike metadata, install-time execution, second-stage loaders, remote access, and credential theft. The payloads searched developer and AI-tool configurations including VS Code, Windsurf, Cursor, AWS, Gemini, Claude, SSH, and shell files.
Bad Epoll Linux kernel flaw lets unprivileged users gain root on Linux and Android
The newly disclosed CVE-2026-46242 is a use-after-free in Linux epoll that affects Linux desktops, servers, and Android; a researcher built a kernelCTF exploit that reportedly reaches root about 99% of the time on tested systems, and a fix is available.
Claude Fable 5 re-release scores 54.8% Pass@1 on Mercor APEX-SWE
Mercor says the July re-release of Fable 5 scored 54.8% Pass@1 on APEX-SWE, below its June 65.5% result but still above Opus 4.8 at 45.3%.
TrendForce expects a major 2027 HBM4 contract-price hike
TrendForce says 2027 HBM4 supply negotiations are underway and suppliers are expected to push substantial price hikes because AI infrastructure demand is crowding DRAM capacity and next-generation HBM manufacturing costs are rising.
RunZero discloses mostly unpatched FatFs flaws in embedded devices
RunZero disclosed seven FatFs vulnerabilities affecting a filesystem library used across security cameras, drones, industrial controllers, hardware wallets, RTOS platforms, and firmware updaters. Malformed USB, SD, or update images can cause memory corruption, data leakage, crashes, and in some cases code execution; PoC material is public.
Tools & repos
5 selected
Antigravity CLI
Terminal-native AI coding agent that brings Antigravity 2.0-style multi-step reasoning, multi-file editing, tool calling, command execution, persistent history, SSH-friendly auth, and session export into a TUI. Useful for keyboard-first and remote workflows, but it should be governed like any code-executing agent.
openai/codex-plugin-cc
OpenAI's Claude Code plugin lets users call Codex from inside Claude Code to review code or delegate tasks, a useful bridge for teams comparing or combining coding agents in one workflow.
Osloq
Osloq is an AI agent for reproducing GitHub issues, aimed at turning vague bug reports into repeatable cases before maintainers spend engineering time debugging them.
Taste Skill v2
Open-source frontend skill files for Cursor, Claude Code, Codex, Gemini CLI, v0, Lovable, OpenCode, and other agents. The v2 ruleset reads the brief, chooses an appropriate design direction or design system, audits existing UI, and bans generic AI-generated frontend patterns.
Glaze by Raycast
Raycast's Glaze lets users create small Mac apps by chatting with AI, a fit for lightweight internal tools and personal automations where a full product build would be overkill.
Blogs worth your time
4 reads
Qwen3-Omni serving in vLLM-Omni
The vLLM team breaks Qwen3-Omni serving into a Thinker, Talker, and Code2Wav pipeline, then shows how batching, CUDA graphs, async handoffs, and speech-stage replicas reduce first-audio latency from about 6s to about 0.6s and lift throughput roughly 5.4x under load.
Modern GPU Programming For MLSys
Tianqi Chen's open book builds a practical path from GPU hardware mental models to Blackwell-era kernels, using GEMM and FlashAttention examples to teach layouts, TMA, tensor cores, barriers, persistent scheduling, and warp specialization.
Open sharded inference of a 229B MoE over the public internet
c0mpute's technical report runs MiniMax-M2.5 across five consumer RTX 5090s in five European countries, measuring 12.6 tok/s interactive and 194 tok/s batched with cryptographic receipts on every request. The useful takeaway is where decentralized inference breaks: host CPUs and speculative decoding over high-latency links, not only raw GPU capacity.
Google DeepMind and A24 announce first-of-its-kind research partnership
Google DeepMind and A24 announced a multi-project research collaboration, with Google also investing in A24, to test AI-assisted creative workflows directly with filmmakers. There are no technical outputs yet, but it is a useful example of frontier labs embedding inside vertical workflows instead of shipping generic tools in isolation.
Funding & acquisitions
1 moves
Bell, Cohere, Hypertec and BUZZ HPC
The companies announced a Canadian sovereign AI infrastructure effort combining Bell's national platform, Cohere's enterprise AI, Hypertec's Canadian-built GPU servers, and BUZZ HPC's AI factory and sovereign AI cloud.
Bengaluru radar
1 events
Bangalore Tech Mixer and Social (Tech / AI / Data / IT)
After-work networking for Bangalore software, data, AI, cybersecurity, product, and startup people at We:Neighborhood. Useful if you want local hiring, referral, and peer-network access rather than a formal talk.