July 28, 2026
Today's top tech stories, deduped across the newsletters I read and briefly summarized. Click a source to open the original article.
AI
Anthropic stakes out its position on open-weight models
Anthropic has published its official stance on open-weight AI models, laying out how it weighs openness against safety risks. The post was the most-discussed story on Hacker News this week.
NVIDIA Ising automates quantum computer calibration
NVIDIA has released Ising Calibration, an open source vision-language model (VLM) that reads diagnostic output from quantum processors and calibrates them automatically using enhanced in-context learning. The approach could cut down significantly on manual quantum hardware calibration work.
Six agent harness capabilities for better model performance
NVIDIA argues that the agent harness — the architecture around the model that renders context, executes tools, and manages state — often matters more for performance than which model you pick. The post lists six concrete harness capabilities developers should prioritize to get more out of AI agents.
Moonshot AI releases weights for 2.8 trillion-parameter Kimi K3
As promised earlier this month, Moonshot AI has released the weights for Kimi K3, a 2.8 trillion-parameter model whose weight files alone run to a hefty 1.56TB. Simon Willison calls it one of the best open-weight models released so far.
An updated guide to which AI to use for what
Ethan Mollick has updated his widely read guide on which AI models are best suited for which tasks, and Simon Willison notes how interesting it is to watch the recommendations shift over time as the models evolve.
Opus 5 benchmarked on SlopCodeBench
A new benchmark report tests Anthropic's Opus 5 on SlopCodeBench, a suite designed to expose coding models that produce "slop" — poorly structured or needlessly complex code. The report was among the most popular posts on Hacker News this week.
Security
Inside the relay market powering LLM token reselling and fraud
An investigative piece by Matt Lenhard uncovers the market that has sprung up around reselling discounted LLM tokens, where operators pool API keys from various sources to offer cut-rate access — often funded by stolen or abused accounts.
Shadow AI agents are multiplying across the enterprise
AI agents are popping up across more and more enterprise platforms, often invisible to IT and security teams. Nudge Security explains how organizations can discover, assess, and govern these agents before unmanaged permissions and autonomous actions turn into security risk.
Attackers exploit FastJson zero-day to hit US firms
Hackers are actively exploiting a vulnerability in the popular open-source Java library FastJson that allows remote code execution with no user interaction and no elevated privileges required. The attacks are currently targeting US firms.
Arista patches actively exploited VeloCloud Orchestrator zero-day
Arista has patched a maximum-severity command injection vulnerability in on-premises VeloCloud Orchestrator deployments that is already being actively exploited in attacks.
Researcher gained full control of Volvo/Eicher's fleet platform
A security researcher demonstrated how flaws in Volvo/Eicher's fleet management platform could be exploited to gain control over every user and vehicle connected to the system, including location data and remote control of vehicle functions.
Dev
Watching Go's new garbage collector move through the heap
A detailed writeup shows how Go's new garbage collector behaves differently from the old one as it moves through the heap, with visualizations that make the changes easy to follow.